{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 78, 503, 116 ], "lines": [ { "bbox": [ 106, 78, 505, 98 ], "spans": [ { "bbox": [ 106, 78, 505, 98 ], "score": 1.0, "content": "DEFECT TRANSFER GAN: DIVERSE DEFECT SYNTHE-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 101, 331, 117 ], "spans": [ { "bbox": [ 106, 101, 331, 117 ], "score": 1.0, "content": "SIS FOR DATA AUGMENTATION", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 112, 136, 244, 157 ], "lines": [ { "bbox": [ 113, 136, 201, 147 ], "spans": [ { "bbox": [ 113, 136, 201, 147 ], "score": 1.0, "content": "Anonymous authors", "type": "text" } ], "index": 2 }, { "bbox": [ 112, 146, 245, 159 ], "spans": [ { "bbox": [ 112, 146, 245, 159 ], "score": 1.0, "content": "Paper under double-blind review", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "title", "bbox": [ 278, 187, 333, 199 ], "lines": [ { "bbox": [ 276, 186, 335, 200 ], "spans": [ { "bbox": [ 276, 186, 335, 200 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 143, 212, 468, 376 ], "lines": [ { "bbox": [ 140, 210, 469, 226 ], "spans": [ { "bbox": [ 140, 210, 469, 226 ], "score": 1.0, "content": "Large amounts of data are a common requirement for many deep learning ap-", "type": "text" } ], "index": 5 }, { "bbox": [ 141, 223, 470, 236 ], "spans": [ { "bbox": [ 141, 223, 470, 236 ], "score": 1.0, "content": "proaches. However, data is not always equally available at large scale for all", "type": "text" } ], "index": 6 }, { "bbox": [ 141, 234, 470, 246 ], "spans": [ { "bbox": [ 141, 234, 470, 246 ], "score": 1.0, "content": "classes. For example, on highly optimized production lines, defective samples are", "type": "text" } ], "index": 7 }, { "bbox": [ 141, 244, 470, 257 ], "spans": [ { "bbox": [ 141, 244, 470, 257 ], "score": 1.0, "content": "hardly acquired while non-defective samples come almost for free. The defects", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 255, 469, 268 ], "spans": [ { "bbox": [ 141, 255, 469, 268 ], "score": 1.0, "content": "however often seem to resemble each other, e.g., scratches on different products", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 267, 470, 279 ], "spans": [ { "bbox": [ 141, 267, 470, 279 ], "score": 1.0, "content": "may only differ in few characteristics. In this work, we propose to make use of", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 277, 470, 290 ], "spans": [ { "bbox": [ 141, 277, 470, 290 ], "score": 1.0, "content": "the shared characteristics by transferring a stylized defect-specific content from", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 289, 470, 300 ], "spans": [ { "bbox": [ 141, 289, 470, 300 ], "score": 1.0, "content": "one type of background product to another. Moreover, the stochastic variations of", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 299, 470, 312 ], "spans": [ { "bbox": [ 141, 299, 470, 312 ], "score": 1.0, "content": "the shared characteristics are captured, which also allows generating novel defects", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 311, 470, 323 ], "spans": [ { "bbox": [ 141, 311, 470, 323 ], "score": 1.0, "content": "from random noise. These synthetic defective samples enlarge the dataset and in-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 322, 470, 333 ], "spans": [ { "bbox": [ 141, 322, 470, 333 ], "score": 1.0, "content": "crease the diversity of defects on the target product. Experiments demonstrate that", "type": "text" } ], "index": 15 }, { "bbox": [ 142, 333, 469, 344 ], "spans": [ { "bbox": [ 142, 333, 469, 344 ], "score": 1.0, "content": "our model is able to disentangle the defect-specific content from the background", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 343, 470, 357 ], "spans": [ { "bbox": [ 141, 343, 470, 357 ], "score": 1.0, "content": "of an image without pixel-level labels. We present convincing results on images", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 354, 469, 366 ], "spans": [ { "bbox": [ 141, 354, 469, 366 ], "score": 1.0, "content": "from real industrial production lines. Also, we show consistent gains of using our", "type": "text" } ], "index": 18 }, { "bbox": [ 142, 366, 357, 378 ], "spans": [ { "bbox": [ 142, 366, 357, 378 ], "score": 1.0, "content": "method to enlarge training sets in classification tasks.", "type": "text" } ], "index": 19 } ], "index": 12 }, { "type": "title", "bbox": [ 108, 398, 206, 411 ], "lines": [ { "bbox": [ 105, 397, 208, 414 ], "spans": [ { "bbox": [ 105, 397, 208, 414 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 424, 505, 555 ], "lines": [ { "bbox": [ 105, 424, 505, 436 ], "spans": [ { "bbox": [ 105, 424, 505, 436 ], "score": 1.0, "content": "Automated Visual Inspection (AVI) is vital for quality control in modern production lines. Despite", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 435, 505, 447 ], "spans": [ { "bbox": [ 105, 435, 505, 447 ], "score": 1.0, "content": "the fact that AVI has been studied for decades, it remains a challenging task with many open research", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 445, 505, 459 ], "spans": [ { "bbox": [ 105, 445, 505, 459 ], "score": 1.0, "content": "questions await to be answered. One of the main challenges in data-driven AVI is the acquisition", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 457, 505, 468 ], "spans": [ { "bbox": [ 106, 457, 505, 468 ], "score": 1.0, "content": "of suitable training data. This is for two reasons: First, collecting a vast amount of labelled data is", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 468, 505, 480 ], "spans": [ { "bbox": [ 105, 468, 505, 480 ], "score": 1.0, "content": "usually labor-intensive and time-consuming. In many cases, even experts are required to identify", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 478, 506, 491 ], "spans": [ { "bbox": [ 105, 478, 506, 491 ], "score": 1.0, "content": "where and what to look for. However, the acquired label information is task-specific and cannot", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 489, 505, 502 ], "spans": [ { "bbox": [ 105, 489, 505, 502 ], "score": 1.0, "content": "be reused or transferred to a new task in most cases. Thus, the tedious labelling process must be", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 500, 505, 514 ], "spans": [ { "bbox": [ 105, 500, 505, 514 ], "score": 1.0, "content": "repeated for each new product, even if its defect is similar to other products in people’s eyes. Second,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 510, 506, 525 ], "spans": [ { "bbox": [ 105, 510, 506, 525 ], "score": 1.0, "content": "in real-world scenarios such as highly optimized production lines, a more severe problem emerges:", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 523, 505, 535 ], "spans": [ { "bbox": [ 105, 523, 505, 535 ], "score": 1.0, "content": "data imbalance. Only very few defective parts are produced by design. Moreover, the acquired", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 534, 505, 546 ], "spans": [ { "bbox": [ 105, 534, 505, 546 ], "score": 1.0, "content": "anomaly images from a single product are lacking diversity and may not capture the full defect", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 543, 495, 558 ], "spans": [ { "bbox": [ 105, 543, 495, 558 ], "score": 1.0, "content": "distribution. Training a robust deep neural network model in such conditions is very challenging.", "type": "text" } ], "index": 32 } ], "index": 26.5 }, { "type": "text", "bbox": [ 107, 561, 504, 627 ], "lines": [ { "bbox": [ 106, 560, 506, 574 ], "spans": [ { "bbox": [ 106, 560, 506, 574 ], "score": 1.0, "content": "Since collecting sufficient real-world defective samples is impractical, algorithms to synthesize re-", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 572, 506, 584 ], "spans": [ { "bbox": [ 105, 572, 506, 584 ], "score": 1.0, "content": "quired images became a focus in research. Image synthesis through Generative Adversarial Net-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 583, 506, 595 ], "spans": [ { "bbox": [ 105, 583, 506, 595 ], "score": 1.0, "content": "works (GANs) (Goodfellow et al., 2014) has shown promising performance in recent years. But it", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 595, 505, 606 ], "spans": [ { "bbox": [ 106, 595, 505, 606 ], "score": 1.0, "content": "also requires large amounts of balanced data which are not available in most industrial use cases, in", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 605, 506, 617 ], "spans": [ { "bbox": [ 105, 605, 506, 617 ], "score": 1.0, "content": "particular for irregular defect patterns and large variation. Therefore, GANs tend to overfit to the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 616, 383, 628 ], "spans": [ { "bbox": [ 105, 616, 383, 628 ], "score": 1.0, "content": "training examples when trained with little data (Karras et al., 2020a).", "type": "text" } ], "index": 38 } ], "index": 35.5 }, { "type": "text", "bbox": [ 107, 633, 505, 732 ], "lines": [ { "bbox": [ 105, 633, 505, 645 ], "spans": [ { "bbox": [ 105, 633, 505, 645 ], "score": 1.0, "content": "In this work, we tackle these issues by exploiting cross-domain information: we first define two", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 643, 505, 657 ], "spans": [ { "bbox": [ 105, 643, 505, 657 ], "score": 1.0, "content": "sets of domains—foreground domains and background domains. The foreground domain describes", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 655, 506, 668 ], "spans": [ { "bbox": [ 104, 655, 506, 668 ], "score": 1.0, "content": "a set of images that contains a specific foreground content to be grouped into a distinctive category,", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 666, 505, 678 ], "spans": [ { "bbox": [ 106, 666, 505, 678 ], "score": 1.0, "content": "and each content has a different style. The background domain instead is considered as a group of", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 677, 506, 690 ], "spans": [ { "bbox": [ 105, 677, 506, 690 ], "score": 1.0, "content": "images that shares similar structural appearance over the whole image. For example, we can set", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "foreground domains as defect types and background domains as product types while the styles of", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "defects indicate their artistic looks such as light or heavy strokes. Building upon StarGAN v2 (Choi", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "et al., 2020), the concept underlying this work is to transfer and generate foreground contents with", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 720, 434, 733 ], "spans": [ { "bbox": [ 105, 720, 434, 733 ], "score": 1.0, "content": "a variety of styles across different background domains, as illustrated in Figure 1.", "type": "text" } ], "index": 47 } ], "index": 43 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 303, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 308, 761 ], "spans": [ { "bbox": [ 302, 751, 308, 761 ], "score": 1.0, "content": "1", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 78, 503, 116 ], "lines": [ { "bbox": [ 106, 78, 505, 98 ], "spans": [ { "bbox": [ 106, 78, 505, 98 ], "score": 1.0, "content": "DEFECT TRANSFER GAN: DIVERSE DEFECT SYNTHE-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 101, 331, 117 ], "spans": [ { "bbox": [ 106, 101, 331, 117 ], "score": 1.0, "content": "SIS FOR DATA AUGMENTATION", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 112, 136, 244, 157 ], "lines": [ { "bbox": [ 113, 136, 201, 147 ], "spans": [ { "bbox": [ 113, 136, 201, 147 ], "score": 1.0, "content": "Anonymous authors", "type": "text" } ], "index": 2 }, { "bbox": [ 112, 146, 245, 159 ], "spans": [ { "bbox": [ 112, 146, 245, 159 ], "score": 1.0, "content": "Paper under double-blind review", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 112, 136, 245, 159 ] }, { "type": "title", "bbox": [ 278, 187, 333, 199 ], "lines": [ { "bbox": [ 276, 186, 335, 200 ], "spans": [ { "bbox": [ 276, 186, 335, 200 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 143, 212, 468, 376 ], "lines": [ { "bbox": [ 140, 210, 469, 226 ], "spans": [ { "bbox": [ 140, 210, 469, 226 ], "score": 1.0, "content": "Large amounts of data are a common requirement for many deep learning ap-", "type": "text" } ], "index": 5 }, { "bbox": [ 141, 223, 470, 236 ], "spans": [ { "bbox": [ 141, 223, 470, 236 ], "score": 1.0, "content": "proaches. However, data is not always equally available at large scale for all", "type": "text" } ], "index": 6 }, { "bbox": [ 141, 234, 470, 246 ], "spans": [ { "bbox": [ 141, 234, 470, 246 ], "score": 1.0, "content": "classes. For example, on highly optimized production lines, defective samples are", "type": "text" } ], "index": 7 }, { "bbox": [ 141, 244, 470, 257 ], "spans": [ { "bbox": [ 141, 244, 470, 257 ], "score": 1.0, "content": "hardly acquired while non-defective samples come almost for free. The defects", "type": "text" } ], "index": 8 }, { "bbox": [ 141, 255, 469, 268 ], "spans": [ { "bbox": [ 141, 255, 469, 268 ], "score": 1.0, "content": "however often seem to resemble each other, e.g., scratches on different products", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 267, 470, 279 ], "spans": [ { "bbox": [ 141, 267, 470, 279 ], "score": 1.0, "content": "may only differ in few characteristics. In this work, we propose to make use of", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 277, 470, 290 ], "spans": [ { "bbox": [ 141, 277, 470, 290 ], "score": 1.0, "content": "the shared characteristics by transferring a stylized defect-specific content from", "type": "text" } ], "index": 11 }, { "bbox": [ 141, 289, 470, 300 ], "spans": [ { "bbox": [ 141, 289, 470, 300 ], "score": 1.0, "content": "one type of background product to another. Moreover, the stochastic variations of", "type": "text" } ], "index": 12 }, { "bbox": [ 141, 299, 470, 312 ], "spans": [ { "bbox": [ 141, 299, 470, 312 ], "score": 1.0, "content": "the shared characteristics are captured, which also allows generating novel defects", "type": "text" } ], "index": 13 }, { "bbox": [ 141, 311, 470, 323 ], "spans": [ { "bbox": [ 141, 311, 470, 323 ], "score": 1.0, "content": "from random noise. These synthetic defective samples enlarge the dataset and in-", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 322, 470, 333 ], "spans": [ { "bbox": [ 141, 322, 470, 333 ], "score": 1.0, "content": "crease the diversity of defects on the target product. Experiments demonstrate that", "type": "text" } ], "index": 15 }, { "bbox": [ 142, 333, 469, 344 ], "spans": [ { "bbox": [ 142, 333, 469, 344 ], "score": 1.0, "content": "our model is able to disentangle the defect-specific content from the background", "type": "text" } ], "index": 16 }, { "bbox": [ 141, 343, 470, 357 ], "spans": [ { "bbox": [ 141, 343, 470, 357 ], "score": 1.0, "content": "of an image without pixel-level labels. We present convincing results on images", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 354, 469, 366 ], "spans": [ { "bbox": [ 141, 354, 469, 366 ], "score": 1.0, "content": "from real industrial production lines. Also, we show consistent gains of using our", "type": "text" } ], "index": 18 }, { "bbox": [ 142, 366, 357, 378 ], "spans": [ { "bbox": [ 142, 366, 357, 378 ], "score": 1.0, "content": "method to enlarge training sets in classification tasks.", "type": "text" } ], "index": 19 } ], "index": 12, "bbox_fs": [ 140, 210, 470, 378 ] }, { "type": "title", "bbox": [ 108, 398, 206, 411 ], "lines": [ { "bbox": [ 105, 397, 208, 414 ], "spans": [ { "bbox": [ 105, 397, 208, 414 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 424, 505, 555 ], "lines": [ { "bbox": [ 105, 424, 505, 436 ], "spans": [ { "bbox": [ 105, 424, 505, 436 ], "score": 1.0, "content": "Automated Visual Inspection (AVI) is vital for quality control in modern production lines. Despite", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 435, 505, 447 ], "spans": [ { "bbox": [ 105, 435, 505, 447 ], "score": 1.0, "content": "the fact that AVI has been studied for decades, it remains a challenging task with many open research", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 445, 505, 459 ], "spans": [ { "bbox": [ 105, 445, 505, 459 ], "score": 1.0, "content": "questions await to be answered. One of the main challenges in data-driven AVI is the acquisition", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 457, 505, 468 ], "spans": [ { "bbox": [ 106, 457, 505, 468 ], "score": 1.0, "content": "of suitable training data. This is for two reasons: First, collecting a vast amount of labelled data is", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 468, 505, 480 ], "spans": [ { "bbox": [ 105, 468, 505, 480 ], "score": 1.0, "content": "usually labor-intensive and time-consuming. In many cases, even experts are required to identify", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 478, 506, 491 ], "spans": [ { "bbox": [ 105, 478, 506, 491 ], "score": 1.0, "content": "where and what to look for. However, the acquired label information is task-specific and cannot", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 489, 505, 502 ], "spans": [ { "bbox": [ 105, 489, 505, 502 ], "score": 1.0, "content": "be reused or transferred to a new task in most cases. Thus, the tedious labelling process must be", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 500, 505, 514 ], "spans": [ { "bbox": [ 105, 500, 505, 514 ], "score": 1.0, "content": "repeated for each new product, even if its defect is similar to other products in people’s eyes. Second,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 510, 506, 525 ], "spans": [ { "bbox": [ 105, 510, 506, 525 ], "score": 1.0, "content": "in real-world scenarios such as highly optimized production lines, a more severe problem emerges:", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 523, 505, 535 ], "spans": [ { "bbox": [ 105, 523, 505, 535 ], "score": 1.0, "content": "data imbalance. Only very few defective parts are produced by design. Moreover, the acquired", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 534, 505, 546 ], "spans": [ { "bbox": [ 105, 534, 505, 546 ], "score": 1.0, "content": "anomaly images from a single product are lacking diversity and may not capture the full defect", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 543, 495, 558 ], "spans": [ { "bbox": [ 105, 543, 495, 558 ], "score": 1.0, "content": "distribution. Training a robust deep neural network model in such conditions is very challenging.", "type": "text" } ], "index": 32 } ], "index": 26.5, "bbox_fs": [ 105, 424, 506, 558 ] }, { "type": "text", "bbox": [ 107, 561, 504, 627 ], "lines": [ { "bbox": [ 106, 560, 506, 574 ], "spans": [ { "bbox": [ 106, 560, 506, 574 ], "score": 1.0, "content": "Since collecting sufficient real-world defective samples is impractical, algorithms to synthesize re-", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 572, 506, 584 ], "spans": [ { "bbox": [ 105, 572, 506, 584 ], "score": 1.0, "content": "quired images became a focus in research. Image synthesis through Generative Adversarial Net-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 583, 506, 595 ], "spans": [ { "bbox": [ 105, 583, 506, 595 ], "score": 1.0, "content": "works (GANs) (Goodfellow et al., 2014) has shown promising performance in recent years. But it", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 595, 505, 606 ], "spans": [ { "bbox": [ 106, 595, 505, 606 ], "score": 1.0, "content": "also requires large amounts of balanced data which are not available in most industrial use cases, in", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 605, 506, 617 ], "spans": [ { "bbox": [ 105, 605, 506, 617 ], "score": 1.0, "content": "particular for irregular defect patterns and large variation. Therefore, GANs tend to overfit to the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 616, 383, 628 ], "spans": [ { "bbox": [ 105, 616, 383, 628 ], "score": 1.0, "content": "training examples when trained with little data (Karras et al., 2020a).", "type": "text" } ], "index": 38 } ], "index": 35.5, "bbox_fs": [ 105, 560, 506, 628 ] }, { "type": "text", "bbox": [ 107, 633, 505, 732 ], "lines": [ { "bbox": [ 105, 633, 505, 645 ], "spans": [ { "bbox": [ 105, 633, 505, 645 ], "score": 1.0, "content": "In this work, we tackle these issues by exploiting cross-domain information: we first define two", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 643, 505, 657 ], "spans": [ { "bbox": [ 105, 643, 505, 657 ], "score": 1.0, "content": "sets of domains—foreground domains and background domains. The foreground domain describes", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 655, 506, 668 ], "spans": [ { "bbox": [ 104, 655, 506, 668 ], "score": 1.0, "content": "a set of images that contains a specific foreground content to be grouped into a distinctive category,", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 666, 505, 678 ], "spans": [ { "bbox": [ 106, 666, 505, 678 ], "score": 1.0, "content": "and each content has a different style. The background domain instead is considered as a group of", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 677, 506, 690 ], "spans": [ { "bbox": [ 105, 677, 506, 690 ], "score": 1.0, "content": "images that shares similar structural appearance over the whole image. For example, we can set", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "foreground domains as defect types and background domains as product types while the styles of", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "defects indicate their artistic looks such as light or heavy strokes. Building upon StarGAN v2 (Choi", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "et al., 2020), the concept underlying this work is to transfer and generate foreground contents with", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 720, 434, 733 ], "spans": [ { "bbox": [ 105, 720, 434, 733 ], "score": 1.0, "content": "a variety of styles across different background domains, as illustrated in Figure 1.", "type": "text" } ], "index": 47 } ], "index": 43, "bbox_fs": [ 104, 633, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 158, 82, 450, 188 ], "blocks": [ { "type": "image_body", "bbox": [ 158, 82, 450, 188 ], "group_id": 0, "lines": [ { "bbox": [ 158, 82, 450, 188 ], "spans": [ { "bbox": [ 158, 82, 450, 188 ], "score": 0.972, "type": "image", "image_path": "7375134a93e4d6c487812a69c420b54b7f33ccdc499b7470357d6f8e1d3d2531.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 158, 82, 450, 117.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 158, 117.33333333333334, 450, 152.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 158, 152.66666666666669, 450, 188.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 201, 505, 224 ], "group_id": 0, "lines": [ { "bbox": [ 106, 201, 505, 214 ], "spans": [ { "bbox": [ 106, 201, 505, 214 ], "score": 1.0, "content": "Figure 1: The underlying concept of DT-GAN is to transfer and generate foreground contents with", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 212, 469, 225 ], "spans": [ { "bbox": [ 105, 212, 469, 225 ], "score": 1.0, "content": "a variety of artistic styles (e.g., light / heavy strokes) across different background domains.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "text", "bbox": [ 107, 246, 505, 345 ], "lines": [ { "bbox": [ 106, 246, 505, 259 ], "spans": [ { "bbox": [ 106, 246, 505, 259 ], "score": 1.0, "content": "The contributions of this work are three-fold: First, we introduce Defect Transfer GAN (DT-GAN),", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 258, 504, 270 ], "spans": [ { "bbox": [ 105, 258, 504, 270 ], "score": 1.0, "content": "a model that learns transferring existing foreground content and generating novel contents onto dif-", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 269, 504, 281 ], "spans": [ { "bbox": [ 106, 269, 504, 281 ], "score": 1.0, "content": "ferent backgrounds at the same time. In the real-world scenario, it allows defect inspection networks", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 280, 504, 291 ], "spans": [ { "bbox": [ 106, 280, 504, 291 ], "score": 1.0, "content": "to learn from a variety of synthetic defective images by composing the foreground defects together", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 290, 505, 303 ], "spans": [ { "bbox": [ 105, 290, 505, 303 ], "score": 1.0, "content": "with various non-defective images from different products. Second, DT-GAN is able to disentangle", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 301, 505, 314 ], "spans": [ { "bbox": [ 105, 301, 505, 314 ], "score": 1.0, "content": "the foreground defect-specific content and the defect-irrelevant background in a weakly-supervised", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 312, 505, 325 ], "spans": [ { "bbox": [ 105, 312, 505, 325 ], "score": 1.0, "content": "manner. Third, extensive experiments show that our method can generate diverse and real-looking", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 322, 505, 337 ], "spans": [ { "bbox": [ 105, 322, 505, 337 ], "score": 1.0, "content": "defective samples even for products with only 20 real defective images. These defective images", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 334, 464, 348 ], "spans": [ { "bbox": [ 105, 334, 464, 348 ], "score": 1.0, "content": "generated by DT-GAN boost the performance in defect inspection networks significantly.", "type": "text" } ], "index": 13 } ], "index": 9 }, { "type": "title", "bbox": [ 108, 364, 211, 377 ], "lines": [ { "bbox": [ 104, 362, 213, 379 ], "spans": [ { "bbox": [ 104, 362, 213, 379 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 390, 505, 490 ], "lines": [ { "bbox": [ 105, 389, 505, 403 ], "spans": [ { "bbox": [ 105, 389, 505, 403 ], "score": 1.0, "content": "GANs have shown their power in many computer vision tasks such as image synthesis (Luciˇ c et al., ´", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 401, 505, 415 ], "spans": [ { "bbox": [ 105, 401, 505, 415 ], "score": 1.0, "content": "2019), style translation (Johnson et al., 2016), super-resolution (Ledig et al., 2017), image impaint-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 413, 505, 426 ], "spans": [ { "bbox": [ 106, 413, 505, 426 ], "score": 1.0, "content": "ing (Pathak et al., 2016) and many other applications. To quantify the performance of GANs, visual", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 424, 504, 437 ], "spans": [ { "bbox": [ 106, 424, 504, 437 ], "score": 1.0, "content": "quality and the diversity of generated images are considered as two of the most important criteria.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 434, 505, 447 ], "spans": [ { "bbox": [ 105, 434, 505, 447 ], "score": 1.0, "content": "Recent models address these requirements either by dedicated loss functions (Mao et al., 2019b;", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 445, 505, 459 ], "spans": [ { "bbox": [ 105, 445, 505, 459 ], "score": 1.0, "content": "Yang et al., 2019) or architectural design (Brock et al., 2019). StyleGAN v2 (Karras et al., 2020b),", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 456, 506, 470 ], "spans": [ { "bbox": [ 105, 456, 506, 470 ], "score": 1.0, "content": "the latest state-of-the-art model in image synthesis, introduces stochastic variation in image gener-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 467, 506, 480 ], "spans": [ { "bbox": [ 105, 467, 506, 480 ], "score": 1.0, "content": "ating process by adding per-pixel noise after each convolution. However, it is non-trivial to adapt", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 479, 415, 491 ], "spans": [ { "bbox": [ 106, 479, 415, 491 ], "score": 1.0, "content": "the model to transform given input images due to the design of the generator.", "type": "text" } ], "index": 23 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 495, 505, 583 ], "lines": [ { "bbox": [ 105, 496, 505, 507 ], "spans": [ { "bbox": [ 105, 496, 505, 507 ], "score": 1.0, "content": "In contrast, image-to-image translation methods (Isola et al., 2017) provide a way to recover the", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 506, 505, 519 ], "spans": [ { "bbox": [ 105, 506, 505, 519 ], "score": 1.0, "content": "connection between inputs and the generated images while encouraging diversity. For example, Zhu", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 517, 505, 531 ], "spans": [ { "bbox": [ 105, 517, 505, 531 ], "score": 1.0, "content": "et al. (2017b) and Huang et al. (2018) impose consistent mappings in latent space to achieve the goal.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 528, 505, 541 ], "spans": [ { "bbox": [ 105, 528, 505, 541 ], "score": 1.0, "content": "Some approaches (Ma et al., 2019; Park et al., 2019) use reference images as guidance to generate", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 540, 506, 552 ], "spans": [ { "bbox": [ 105, 540, 506, 552 ], "score": 1.0, "content": "diverse outputs. Mokady et al. (2020) further extends the translation task from styles to contents. It", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 550, 507, 563 ], "spans": [ { "bbox": [ 105, 550, 507, 563 ], "score": 1.0, "content": "learns to identify a specific content in a given input (e.g., a specific pair of glasses) and transfer it", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 561, 506, 574 ], "spans": [ { "bbox": [ 105, 561, 506, 574 ], "score": 1.0, "content": "to the target image. However, aforementioned methods only consider the translation between two", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 572, 359, 584 ], "spans": [ { "bbox": [ 106, 572, 359, 584 ], "score": 1.0, "content": "domains and their extension to multiple domains is non-trivial.", "type": "text" } ], "index": 31 } ], "index": 27.5 }, { "type": "text", "bbox": [ 106, 589, 505, 732 ], "lines": [ { "bbox": [ 106, 588, 505, 601 ], "spans": [ { "bbox": [ 106, 588, 505, 601 ], "score": 1.0, "content": "Surface defect detection is one of the important tasks in real-world industrial manufacturing. It", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 600, 505, 612 ], "spans": [ { "bbox": [ 105, 600, 505, 612 ], "score": 1.0, "content": "aims at identifying and classifying defects with the help of machine vision. Traditional methods", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 610, 506, 623 ], "spans": [ { "bbox": [ 105, 610, 506, 623 ], "score": 1.0, "content": "(Ngan et al., 2011) build models upon hand-crafted feature extractors, which are unstable and out-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 622, 506, 634 ], "spans": [ { "bbox": [ 105, 622, 506, 634 ], "score": 1.0, "content": "performed by deep learning based models. However, the performance and generalization ability of", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 633, 506, 645 ], "spans": [ { "bbox": [ 105, 633, 506, 645 ], "score": 1.0, "content": "deep learning approaches are restricted due to limited number of defective samples in real-world", "type": "text" } ], "index": 36 }, { "bbox": [ 104, 644, 506, 656 ], "spans": [ { "bbox": [ 104, 644, 506, 656 ], "score": 1.0, "content": "scenarios. Data augmentation aims to enrich the training dataset by introducing different kinds of", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 655, 506, 668 ], "spans": [ { "bbox": [ 105, 655, 506, 668 ], "score": 1.0, "content": "invariance for the model to capture. Several recent works (Niu et al., 2020; Zhang et al., 2021)", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 666, 506, 678 ], "spans": [ { "bbox": [ 105, 666, 506, 678 ], "score": 1.0, "content": "have proposed to adopt GANs as a data augmentation method to generate realistic defective sam-", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "ples. Among them, Defect-GAN (Zhang et al., 2021) tries to capture the stochastic variation within", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "defects by mimicking the defacement and restoration processes. However, it still learns a deter-", "type": "text" } ], "index": 41 }, { "bbox": [ 104, 696, 506, 714 ], "spans": [ { "bbox": [ 104, 696, 506, 714 ], "score": 1.0, "content": "ministic mapping between inputs and outputs while DT-GAN achieves multi-modality by varying", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 710, 505, 723 ], "spans": [ { "bbox": [ 106, 710, 505, 723 ], "score": 1.0, "content": "styles. Moreover, our method can generate realistic defects with sophisticated patterns copied from", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 720, 225, 732 ], "spans": [ { "bbox": [ 106, 720, 225, 732 ], "score": 1.0, "content": "real-world defective samples.", "type": "text" } ], "index": 44 } ], "index": 38 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 309, 38 ], "spans": [ { "bbox": [ 106, 26, 309, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 763 ], "spans": [ { "bbox": [ 301, 750, 310, 763 ], "score": 1.0, "content": "2", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 158, 82, 450, 188 ], "blocks": [ { "type": "image_body", "bbox": [ 158, 82, 450, 188 ], "group_id": 0, "lines": [ { "bbox": [ 158, 82, 450, 188 ], "spans": [ { "bbox": [ 158, 82, 450, 188 ], "score": 0.972, "type": "image", "image_path": "7375134a93e4d6c487812a69c420b54b7f33ccdc499b7470357d6f8e1d3d2531.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 158, 82, 450, 117.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 158, 117.33333333333334, 450, 152.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 158, 152.66666666666669, 450, 188.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 201, 505, 224 ], "group_id": 0, "lines": [ { "bbox": [ 106, 201, 505, 214 ], "spans": [ { "bbox": [ 106, 201, 505, 214 ], "score": 1.0, "content": "Figure 1: The underlying concept of DT-GAN is to transfer and generate foreground contents with", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 212, 469, 225 ], "spans": [ { "bbox": [ 105, 212, 469, 225 ], "score": 1.0, "content": "a variety of artistic styles (e.g., light / heavy strokes) across different background domains.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "text", "bbox": [ 107, 246, 505, 345 ], "lines": [ { "bbox": [ 106, 246, 505, 259 ], "spans": [ { "bbox": [ 106, 246, 505, 259 ], "score": 1.0, "content": "The contributions of this work are three-fold: First, we introduce Defect Transfer GAN (DT-GAN),", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 258, 504, 270 ], "spans": [ { "bbox": [ 105, 258, 504, 270 ], "score": 1.0, "content": "a model that learns transferring existing foreground content and generating novel contents onto dif-", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 269, 504, 281 ], "spans": [ { "bbox": [ 106, 269, 504, 281 ], "score": 1.0, "content": "ferent backgrounds at the same time. In the real-world scenario, it allows defect inspection networks", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 280, 504, 291 ], "spans": [ { "bbox": [ 106, 280, 504, 291 ], "score": 1.0, "content": "to learn from a variety of synthetic defective images by composing the foreground defects together", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 290, 505, 303 ], "spans": [ { "bbox": [ 105, 290, 505, 303 ], "score": 1.0, "content": "with various non-defective images from different products. Second, DT-GAN is able to disentangle", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 301, 505, 314 ], "spans": [ { "bbox": [ 105, 301, 505, 314 ], "score": 1.0, "content": "the foreground defect-specific content and the defect-irrelevant background in a weakly-supervised", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 312, 505, 325 ], "spans": [ { "bbox": [ 105, 312, 505, 325 ], "score": 1.0, "content": "manner. Third, extensive experiments show that our method can generate diverse and real-looking", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 322, 505, 337 ], "spans": [ { "bbox": [ 105, 322, 505, 337 ], "score": 1.0, "content": "defective samples even for products with only 20 real defective images. These defective images", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 334, 464, 348 ], "spans": [ { "bbox": [ 105, 334, 464, 348 ], "score": 1.0, "content": "generated by DT-GAN boost the performance in defect inspection networks significantly.", "type": "text" } ], "index": 13 } ], "index": 9, "bbox_fs": [ 105, 246, 505, 348 ] }, { "type": "title", "bbox": [ 108, 364, 211, 377 ], "lines": [ { "bbox": [ 104, 362, 213, 379 ], "spans": [ { "bbox": [ 104, 362, 213, 379 ], "score": 1.0, "content": "2 RELATED WORK", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 106, 390, 505, 490 ], "lines": [ { "bbox": [ 105, 389, 505, 403 ], "spans": [ { "bbox": [ 105, 389, 505, 403 ], "score": 1.0, "content": "GANs have shown their power in many computer vision tasks such as image synthesis (Luciˇ c et al., ´", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 401, 505, 415 ], "spans": [ { "bbox": [ 105, 401, 505, 415 ], "score": 1.0, "content": "2019), style translation (Johnson et al., 2016), super-resolution (Ledig et al., 2017), image impaint-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 413, 505, 426 ], "spans": [ { "bbox": [ 106, 413, 505, 426 ], "score": 1.0, "content": "ing (Pathak et al., 2016) and many other applications. To quantify the performance of GANs, visual", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 424, 504, 437 ], "spans": [ { "bbox": [ 106, 424, 504, 437 ], "score": 1.0, "content": "quality and the diversity of generated images are considered as two of the most important criteria.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 434, 505, 447 ], "spans": [ { "bbox": [ 105, 434, 505, 447 ], "score": 1.0, "content": "Recent models address these requirements either by dedicated loss functions (Mao et al., 2019b;", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 445, 505, 459 ], "spans": [ { "bbox": [ 105, 445, 505, 459 ], "score": 1.0, "content": "Yang et al., 2019) or architectural design (Brock et al., 2019). StyleGAN v2 (Karras et al., 2020b),", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 456, 506, 470 ], "spans": [ { "bbox": [ 105, 456, 506, 470 ], "score": 1.0, "content": "the latest state-of-the-art model in image synthesis, introduces stochastic variation in image gener-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 467, 506, 480 ], "spans": [ { "bbox": [ 105, 467, 506, 480 ], "score": 1.0, "content": "ating process by adding per-pixel noise after each convolution. However, it is non-trivial to adapt", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 479, 415, 491 ], "spans": [ { "bbox": [ 106, 479, 415, 491 ], "score": 1.0, "content": "the model to transform given input images due to the design of the generator.", "type": "text" } ], "index": 23 } ], "index": 19, "bbox_fs": [ 105, 389, 506, 491 ] }, { "type": "text", "bbox": [ 107, 495, 505, 583 ], "lines": [ { "bbox": [ 105, 496, 505, 507 ], "spans": [ { "bbox": [ 105, 496, 505, 507 ], "score": 1.0, "content": "In contrast, image-to-image translation methods (Isola et al., 2017) provide a way to recover the", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 506, 505, 519 ], "spans": [ { "bbox": [ 105, 506, 505, 519 ], "score": 1.0, "content": "connection between inputs and the generated images while encouraging diversity. For example, Zhu", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 517, 505, 531 ], "spans": [ { "bbox": [ 105, 517, 505, 531 ], "score": 1.0, "content": "et al. (2017b) and Huang et al. (2018) impose consistent mappings in latent space to achieve the goal.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 528, 505, 541 ], "spans": [ { "bbox": [ 105, 528, 505, 541 ], "score": 1.0, "content": "Some approaches (Ma et al., 2019; Park et al., 2019) use reference images as guidance to generate", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 540, 506, 552 ], "spans": [ { "bbox": [ 105, 540, 506, 552 ], "score": 1.0, "content": "diverse outputs. Mokady et al. (2020) further extends the translation task from styles to contents. It", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 550, 507, 563 ], "spans": [ { "bbox": [ 105, 550, 507, 563 ], "score": 1.0, "content": "learns to identify a specific content in a given input (e.g., a specific pair of glasses) and transfer it", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 561, 506, 574 ], "spans": [ { "bbox": [ 105, 561, 506, 574 ], "score": 1.0, "content": "to the target image. However, aforementioned methods only consider the translation between two", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 572, 359, 584 ], "spans": [ { "bbox": [ 106, 572, 359, 584 ], "score": 1.0, "content": "domains and their extension to multiple domains is non-trivial.", "type": "text" } ], "index": 31 } ], "index": 27.5, "bbox_fs": [ 105, 496, 507, 584 ] }, { "type": "text", "bbox": [ 106, 589, 505, 732 ], "lines": [ { "bbox": [ 106, 588, 505, 601 ], "spans": [ { "bbox": [ 106, 588, 505, 601 ], "score": 1.0, "content": "Surface defect detection is one of the important tasks in real-world industrial manufacturing. It", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 600, 505, 612 ], "spans": [ { "bbox": [ 105, 600, 505, 612 ], "score": 1.0, "content": "aims at identifying and classifying defects with the help of machine vision. Traditional methods", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 610, 506, 623 ], "spans": [ { "bbox": [ 105, 610, 506, 623 ], "score": 1.0, "content": "(Ngan et al., 2011) build models upon hand-crafted feature extractors, which are unstable and out-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 622, 506, 634 ], "spans": [ { "bbox": [ 105, 622, 506, 634 ], "score": 1.0, "content": "performed by deep learning based models. However, the performance and generalization ability of", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 633, 506, 645 ], "spans": [ { "bbox": [ 105, 633, 506, 645 ], "score": 1.0, "content": "deep learning approaches are restricted due to limited number of defective samples in real-world", "type": "text" } ], "index": 36 }, { "bbox": [ 104, 644, 506, 656 ], "spans": [ { "bbox": [ 104, 644, 506, 656 ], "score": 1.0, "content": "scenarios. Data augmentation aims to enrich the training dataset by introducing different kinds of", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 655, 506, 668 ], "spans": [ { "bbox": [ 105, 655, 506, 668 ], "score": 1.0, "content": "invariance for the model to capture. Several recent works (Niu et al., 2020; Zhang et al., 2021)", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 666, 506, 678 ], "spans": [ { "bbox": [ 105, 666, 506, 678 ], "score": 1.0, "content": "have proposed to adopt GANs as a data augmentation method to generate realistic defective sam-", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "ples. Among them, Defect-GAN (Zhang et al., 2021) tries to capture the stochastic variation within", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "defects by mimicking the defacement and restoration processes. However, it still learns a deter-", "type": "text" } ], "index": 41 }, { "bbox": [ 104, 696, 506, 714 ], "spans": [ { "bbox": [ 104, 696, 506, 714 ], "score": 1.0, "content": "ministic mapping between inputs and outputs while DT-GAN achieves multi-modality by varying", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 710, 505, 723 ], "spans": [ { "bbox": [ 106, 710, 505, 723 ], "score": 1.0, "content": "styles. Moreover, our method can generate realistic defects with sophisticated patterns copied from", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 720, 225, 732 ], "spans": [ { "bbox": [ 106, 720, 225, 732 ], "score": 1.0, "content": "real-world defective samples.", "type": "text" } ], "index": 44 } ], "index": 38, "bbox_fs": [ 104, 588, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 78, 503, 216 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 78, 503, 216 ], "group_id": 0, "lines": [ { "bbox": [ 106, 78, 503, 216 ], "spans": [ { "bbox": [ 106, 78, 503, 216 ], "score": 0.968, "type": "image", "image_path": "83af523dcfeb8e486bbe482fe2fcb6e531192ec8f0ab542aeb5d211eff025025.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 78, 503, 124.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 124.0, 503, 170.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 170.0, 503, 216.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 209, 217, 401, 229 ], "group_id": 0, "lines": [ { "bbox": [ 208, 216, 402, 230 ], "spans": [ { "bbox": [ 208, 216, 402, 230 ], "score": 1.0, "content": "Figure 2: Overview of all modules in DT-GAN.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "title", "bbox": [ 107, 251, 209, 264 ], "lines": [ { "bbox": [ 104, 249, 211, 267 ], "spans": [ { "bbox": [ 104, 249, 211, 267 ], "score": 1.0, "content": "3 METHODOLOGY", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 278, 505, 334 ], "lines": [ { "bbox": [ 106, 279, 504, 291 ], "spans": [ { "bbox": [ 106, 279, 504, 291 ], "score": 1.0, "content": "Our primary aim is to perform unpaired image-to-image translation across multiple foreground do-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 288, 505, 303 ], "spans": [ { "bbox": [ 105, 288, 505, 303 ], "score": 1.0, "content": "mains within a single model. In our use case, the foreground domains refer to the defect types,", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 300, 505, 314 ], "spans": [ { "bbox": [ 105, 300, 505, 314 ], "score": 1.0, "content": "which means we want to achieve translations between different types of defects while the back-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 311, 506, 325 ], "spans": [ { "bbox": [ 105, 311, 506, 325 ], "score": 1.0, "content": "ground remains unaffected. We assume that there is always an adequate amount of normal samples", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 322, 441, 336 ], "spans": [ { "bbox": [ 105, 322, 441, 336 ], "score": 1.0, "content": "(e.g., non-defective) available, while anomaly samples are rare and hard to acquire.", "type": "text" } ], "index": 9 } ], "index": 7 }, { "type": "title", "bbox": [ 107, 352, 237, 363 ], "lines": [ { "bbox": [ 106, 352, 238, 363 ], "spans": [ { "bbox": [ 106, 352, 238, 363 ], "score": 1.0, "content": "3.1 PROPOSED FRAMEWORK", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 107, 373, 505, 439 ], "lines": [ { "bbox": [ 105, 372, 506, 386 ], "spans": [ { "bbox": [ 105, 372, 506, 386 ], "score": 1.0, "content": "Our framework builds on StarGAN v2, a multimodal image-to-image translation model. Given an", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 385, 504, 397 ], "spans": [ { "bbox": [ 106, 385, 158, 397 ], "score": 1.0, "content": "input image", "type": "text" }, { "bbox": [ 159, 385, 192, 396 ], "score": 0.89, "content": "\\textbf { x } \\in { \\mathcal { X } }", "type": "inline_equation" }, { "bbox": [ 192, 385, 297, 397 ], "score": 1.0, "content": "and an arbitrary domain", "type": "text" }, { "bbox": [ 297, 385, 328, 397 ], "score": 0.91, "content": "y \\in \\mathcal { V }", "type": "inline_equation" }, { "bbox": [ 329, 385, 504, 397 ], "score": 1.0, "content": ", StarGAN v2 generates a domain specific", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 396, 504, 408 ], "spans": [ { "bbox": [ 106, 396, 480, 408 ], "score": 1.0, "content": "style code in a learned style space and outputs an image that is stylized to fit the domain of", "type": "text" }, { "bbox": [ 480, 398, 487, 407 ], "score": 0.77, "content": "y", "type": "inline_equation" }, { "bbox": [ 487, 396, 504, 408 ], "score": 1.0, "content": ". Its", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 407, 505, 420 ], "spans": [ { "bbox": [ 105, 407, 505, 420 ], "score": 1.0, "content": "network architecture consists of four modules: a generator, a mapping network, a style encoder", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 417, 505, 430 ], "spans": [ { "bbox": [ 105, 417, 505, 430 ], "score": 1.0, "content": "and a discriminator. We modify and extend all four modules (see Figure 2) and describe the key", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 428, 231, 441 ], "spans": [ { "bbox": [ 106, 428, 231, 441 ], "score": 1.0, "content": "differences in details as below.", "type": "text" } ], "index": 16 } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 445, 505, 534 ], "lines": [ { "bbox": [ 105, 444, 505, 459 ], "spans": [ { "bbox": [ 105, 444, 303, 459 ], "score": 1.0, "content": "Style-Content Separation. Given a latent code", "type": "text" }, { "bbox": [ 303, 448, 310, 456 ], "score": 0.59, "content": "\\mathbf { z }", "type": "inline_equation" }, { "bbox": [ 311, 444, 369, 459 ], "score": 1.0, "content": "and a domain", "type": "text" }, { "bbox": [ 369, 448, 376, 457 ], "score": 0.8, "content": "y", "type": "inline_equation" }, { "bbox": [ 376, 444, 469, 459 ], "score": 1.0, "content": ", the mapping network", "type": "text" }, { "bbox": [ 469, 446, 481, 456 ], "score": 0.72, "content": "M", "type": "inline_equation" }, { "bbox": [ 482, 444, 505, 459 ], "score": 1.0, "content": "(Fig-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 456, 506, 470 ], "spans": [ { "bbox": [ 105, 456, 235, 470 ], "score": 1.0, "content": "ure 2(b)) generates a style code", "type": "text" }, { "bbox": [ 236, 456, 283, 468 ], "score": 0.94, "content": "\\mathbf { s } = M _ { y } ( \\mathbf { z } )", "type": "inline_equation" }, { "bbox": [ 284, 456, 408, 470 ], "score": 1.0, "content": "and a domain specific content", "type": "text" }, { "bbox": [ 408, 457, 456, 469 ], "score": 0.93, "content": "\\mathbf { c } = M _ { y } ( \\mathbf { z } )", "type": "inline_equation" }, { "bbox": [ 456, 456, 506, 470 ], "score": 1.0, "content": "in different", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 467, 505, 480 ], "spans": [ { "bbox": [ 105, 467, 255, 480 ], "score": 1.0, "content": "branches. It is worth mentioning that", "type": "text" }, { "bbox": [ 256, 468, 271, 480 ], "score": 0.9, "content": "M _ { y }", "type": "inline_equation" }, { "bbox": [ 272, 467, 375, 480 ], "score": 1.0, "content": "here denotes an output of", "type": "text" }, { "bbox": [ 376, 468, 388, 478 ], "score": 0.77, "content": "M", "type": "inline_equation" }, { "bbox": [ 388, 467, 505, 480 ], "score": 1.0, "content": "corresponding to the domain", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 478, 506, 492 ], "spans": [ { "bbox": [ 106, 480, 113, 490 ], "score": 0.68, "content": "y", "type": "inline_equation" }, { "bbox": [ 113, 478, 506, 492 ], "score": 1.0, "content": ". This feature allows our method to separately model the structural appearance (i.e. content) and", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 506, 502 ], "score": 1.0, "content": "its artistic looks (i.e. style), which is essential because applying different styles to the same content", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 500, 504, 514 ], "spans": [ { "bbox": [ 105, 500, 331, 514 ], "score": 1.0, "content": "enriches the diversity of outputs. By randomly sampling", "type": "text" }, { "bbox": [ 331, 502, 338, 511 ], "score": 0.6, "content": "\\mathbf { z }", "type": "inline_equation" }, { "bbox": [ 338, 500, 497, 514 ], "score": 1.0, "content": "from a standard normal distribution and", "type": "text" }, { "bbox": [ 497, 502, 504, 512 ], "score": 0.76, "content": "y", "type": "inline_equation" } ], "index": 22 }, { "bbox": [ 106, 511, 505, 524 ], "spans": [ { "bbox": [ 106, 511, 262, 524 ], "score": 1.0, "content": "from all available foreground domains,", "type": "text" }, { "bbox": [ 263, 512, 275, 521 ], "score": 0.79, "content": "M", "type": "inline_equation" }, { "bbox": [ 275, 511, 505, 524 ], "score": 1.0, "content": "is able to produce diverse style codes and domain specific", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 524, 145, 534 ], "spans": [ { "bbox": [ 106, 524, 145, 534 ], "score": 1.0, "content": "contents.", "type": "text" } ], "index": 24 } ], "index": 20.5 }, { "type": "text", "bbox": [ 109, 539, 503, 573 ], "lines": [ { "bbox": [ 106, 539, 505, 552 ], "spans": [ { "bbox": [ 106, 539, 160, 552 ], "score": 1.0, "content": "The encoder", "type": "text" }, { "bbox": [ 160, 540, 170, 549 ], "score": 0.78, "content": "E", "type": "inline_equation" }, { "bbox": [ 170, 539, 321, 552 ], "score": 1.0, "content": "(Figure 2(c)) extracts the style code", "type": "text" }, { "bbox": [ 322, 540, 370, 551 ], "score": 0.93, "content": "\\mathbf { s } = E _ { y } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 370, 539, 505, 552 ], "score": 1.0, "content": "and the domain specific content", "type": "text" } ], "index": 25 }, { "bbox": [ 107, 550, 505, 563 ], "spans": [ { "bbox": [ 107, 550, 153, 562 ], "score": 0.93, "content": "\\mathbf { c } = E _ { y } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 153, 550, 240, 563 ], "score": 1.0, "content": "from an given image", "type": "text" }, { "bbox": [ 240, 552, 248, 560 ], "score": 0.55, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 248, 550, 505, 563 ], "score": 1.0, "content": ", which reflect the characteristics of reference images instead of", "type": "text" } ], "index": 26 }, { "bbox": [ 107, 561, 209, 574 ], "spans": [ { "bbox": [ 107, 561, 209, 574 ], "score": 1.0, "content": "randomly sampled noise.", "type": "text" } ], "index": 27 } ], "index": 26 }, { "type": "text", "bbox": [ 106, 578, 505, 732 ], "lines": [ { "bbox": [ 106, 578, 505, 591 ], "spans": [ { "bbox": [ 106, 578, 399, 591 ], "score": 1.0, "content": "Foreground/Background (FG/BG) Disentanglement. The generator", "type": "text" }, { "bbox": [ 399, 579, 408, 588 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 409, 578, 505, 591 ], "score": 1.0, "content": "(Figure 2(a)) translates", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 589, 506, 602 ], "spans": [ { "bbox": [ 106, 589, 258, 602 ], "score": 1.0, "content": "an input image x into an output image", "type": "text" }, { "bbox": [ 258, 589, 299, 601 ], "score": 0.93, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 299, 589, 506, 602 ], "score": 1.0, "content": "according to given domain specific style code es and", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 600, 506, 613 ], "spans": [ { "bbox": [ 106, 600, 347, 613 ], "score": 1.0, "content": "content ec, which are provided either by the mapping network", "type": "text" }, { "bbox": [ 348, 601, 360, 610 ], "score": 0.77, "content": "M", "type": "inline_equation" }, { "bbox": [ 360, 600, 506, 613 ], "score": 1.0, "content": "when generating from random noise", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 612, 505, 623 ], "spans": [ { "bbox": [ 106, 612, 232, 623 ], "score": 1.0, "content": "or by the style-content encoder", "type": "text" }, { "bbox": [ 233, 612, 242, 621 ], "score": 0.8, "content": "E", "type": "inline_equation" }, { "bbox": [ 242, 612, 505, 623 ], "score": 1.0, "content": "when transferring an existing content from a reference image. To", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 621, 505, 635 ], "spans": [ { "bbox": [ 105, 621, 505, 635 ], "score": 1.0, "content": "achieve a FG/BG disentanglement, we split the channels of the three-dimensional feature map", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 632, 506, 645 ], "spans": [ { "bbox": [ 105, 632, 128, 645 ], "score": 1.0, "content": "(i.e.,", "type": "text" }, { "bbox": [ 129, 633, 186, 644 ], "score": 0.91, "content": "H \\times W \\times C", "type": "inline_equation" }, { "bbox": [ 186, 632, 278, 645 ], "score": 1.0, "content": ") at the bottle neck of", "type": "text" }, { "bbox": [ 278, 633, 287, 643 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 288, 632, 506, 645 ], "score": 1.0, "content": "into two parts. The model is then forced to encode", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 643, 505, 657 ], "spans": [ { "bbox": [ 105, 643, 505, 657 ], "score": 1.0, "content": "the background into the first channels and the domain specific content cˆ into the latter channels by", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 654, 506, 668 ], "spans": [ { "bbox": [ 105, 654, 506, 668 ], "score": 1.0, "content": "classification losses as discussed in Section 3.2. cˆ can then be replaced with content ec from the target", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 665, 506, 679 ], "spans": [ { "bbox": [ 105, 665, 506, 679 ], "score": 1.0, "content": "domain. The adaptive instance normalization (AdaIN) (Huang & Belongie, 2017) is then used to", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 677, 505, 690 ], "spans": [ { "bbox": [ 106, 677, 374, 690 ], "score": 1.0, "content": "inject es into ec during the decoding process while the background", "type": "text" }, { "bbox": [ 375, 677, 412, 689 ], "score": 0.92, "content": "B G _ { G } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 412, 677, 505, 690 ], "score": 1.0, "content": "is decoded separately.", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 187, 700 ], "score": 1.0, "content": "StarGAN v2 learns", "type": "text" }, { "bbox": [ 187, 688, 204, 698 ], "score": 0.31, "content": "\\mathbf { F G }", "type": "inline_equation" }, { "bbox": [ 204, 687, 505, 700 ], "score": 1.0, "content": "and BG together which leads to a conditional relationship between both.", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Our disentanglement and separate encoding break this conditioning and therefore enable our method", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 709, 504, 722 ], "spans": [ { "bbox": [ 105, 710, 466, 722 ], "score": 1.0, "content": "to freely combine FG and BG as well as learn the full variation of FG content. Finally,", "type": "text" }, { "bbox": [ 466, 709, 504, 722 ], "score": 0.92, "content": "B G _ { G } ( \\mathbf { x } )", "type": "inline_equation" } ], "index": 40 }, { "bbox": [ 106, 721, 354, 733 ], "spans": [ { "bbox": [ 106, 721, 354, 733 ], "score": 1.0, "content": "and ec are concatenated together and then fused before output.", "type": "text" } ], "index": 41 } ], "index": 34.5 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 78, 503, 216 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 78, 503, 216 ], "group_id": 0, "lines": [ { "bbox": [ 106, 78, 503, 216 ], "spans": [ { "bbox": [ 106, 78, 503, 216 ], "score": 0.968, "type": "image", "image_path": "83af523dcfeb8e486bbe482fe2fcb6e531192ec8f0ab542aeb5d211eff025025.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 78, 503, 124.0 ], "spans": [], "index": 0 }, { "bbox": [ 106, 124.0, 503, 170.0 ], "spans": [], "index": 1 }, { "bbox": [ 106, 170.0, 503, 216.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 209, 217, 401, 229 ], "group_id": 0, "lines": [ { "bbox": [ 208, 216, 402, 230 ], "spans": [ { "bbox": [ 208, 216, 402, 230 ], "score": 1.0, "content": "Figure 2: Overview of all modules in DT-GAN.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "title", "bbox": [ 107, 251, 209, 264 ], "lines": [ { "bbox": [ 104, 249, 211, 267 ], "spans": [ { "bbox": [ 104, 249, 211, 267 ], "score": 1.0, "content": "3 METHODOLOGY", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 278, 505, 334 ], "lines": [ { "bbox": [ 106, 279, 504, 291 ], "spans": [ { "bbox": [ 106, 279, 504, 291 ], "score": 1.0, "content": "Our primary aim is to perform unpaired image-to-image translation across multiple foreground do-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 288, 505, 303 ], "spans": [ { "bbox": [ 105, 288, 505, 303 ], "score": 1.0, "content": "mains within a single model. In our use case, the foreground domains refer to the defect types,", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 300, 505, 314 ], "spans": [ { "bbox": [ 105, 300, 505, 314 ], "score": 1.0, "content": "which means we want to achieve translations between different types of defects while the back-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 311, 506, 325 ], "spans": [ { "bbox": [ 105, 311, 506, 325 ], "score": 1.0, "content": "ground remains unaffected. We assume that there is always an adequate amount of normal samples", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 322, 441, 336 ], "spans": [ { "bbox": [ 105, 322, 441, 336 ], "score": 1.0, "content": "(e.g., non-defective) available, while anomaly samples are rare and hard to acquire.", "type": "text" } ], "index": 9 } ], "index": 7, "bbox_fs": [ 105, 279, 506, 336 ] }, { "type": "title", "bbox": [ 107, 352, 237, 363 ], "lines": [ { "bbox": [ 106, 352, 238, 363 ], "spans": [ { "bbox": [ 106, 352, 238, 363 ], "score": 1.0, "content": "3.1 PROPOSED FRAMEWORK", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "text", "bbox": [ 107, 373, 505, 439 ], "lines": [ { "bbox": [ 105, 372, 506, 386 ], "spans": [ { "bbox": [ 105, 372, 506, 386 ], "score": 1.0, "content": "Our framework builds on StarGAN v2, a multimodal image-to-image translation model. Given an", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 385, 504, 397 ], "spans": [ { "bbox": [ 106, 385, 158, 397 ], "score": 1.0, "content": "input image", "type": "text" }, { "bbox": [ 159, 385, 192, 396 ], "score": 0.89, "content": "\\textbf { x } \\in { \\mathcal { X } }", "type": "inline_equation" }, { "bbox": [ 192, 385, 297, 397 ], "score": 1.0, "content": "and an arbitrary domain", "type": "text" }, { "bbox": [ 297, 385, 328, 397 ], "score": 0.91, "content": "y \\in \\mathcal { V }", "type": "inline_equation" }, { "bbox": [ 329, 385, 504, 397 ], "score": 1.0, "content": ", StarGAN v2 generates a domain specific", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 396, 504, 408 ], "spans": [ { "bbox": [ 106, 396, 480, 408 ], "score": 1.0, "content": "style code in a learned style space and outputs an image that is stylized to fit the domain of", "type": "text" }, { "bbox": [ 480, 398, 487, 407 ], "score": 0.77, "content": "y", "type": "inline_equation" }, { "bbox": [ 487, 396, 504, 408 ], "score": 1.0, "content": ". Its", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 407, 505, 420 ], "spans": [ { "bbox": [ 105, 407, 505, 420 ], "score": 1.0, "content": "network architecture consists of four modules: a generator, a mapping network, a style encoder", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 417, 505, 430 ], "spans": [ { "bbox": [ 105, 417, 505, 430 ], "score": 1.0, "content": "and a discriminator. We modify and extend all four modules (see Figure 2) and describe the key", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 428, 231, 441 ], "spans": [ { "bbox": [ 106, 428, 231, 441 ], "score": 1.0, "content": "differences in details as below.", "type": "text" } ], "index": 16 } ], "index": 13.5, "bbox_fs": [ 105, 372, 506, 441 ] }, { "type": "text", "bbox": [ 106, 445, 505, 534 ], "lines": [ { "bbox": [ 105, 444, 505, 459 ], "spans": [ { "bbox": [ 105, 444, 303, 459 ], "score": 1.0, "content": "Style-Content Separation. Given a latent code", "type": "text" }, { "bbox": [ 303, 448, 310, 456 ], "score": 0.59, "content": "\\mathbf { z }", "type": "inline_equation" }, { "bbox": [ 311, 444, 369, 459 ], "score": 1.0, "content": "and a domain", "type": "text" }, { "bbox": [ 369, 448, 376, 457 ], "score": 0.8, "content": "y", "type": "inline_equation" }, { "bbox": [ 376, 444, 469, 459 ], "score": 1.0, "content": ", the mapping network", "type": "text" }, { "bbox": [ 469, 446, 481, 456 ], "score": 0.72, "content": "M", "type": "inline_equation" }, { "bbox": [ 482, 444, 505, 459 ], "score": 1.0, "content": "(Fig-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 456, 506, 470 ], "spans": [ { "bbox": [ 105, 456, 235, 470 ], "score": 1.0, "content": "ure 2(b)) generates a style code", "type": "text" }, { "bbox": [ 236, 456, 283, 468 ], "score": 0.94, "content": "\\mathbf { s } = M _ { y } ( \\mathbf { z } )", "type": "inline_equation" }, { "bbox": [ 284, 456, 408, 470 ], "score": 1.0, "content": "and a domain specific content", "type": "text" }, { "bbox": [ 408, 457, 456, 469 ], "score": 0.93, "content": "\\mathbf { c } = M _ { y } ( \\mathbf { z } )", "type": "inline_equation" }, { "bbox": [ 456, 456, 506, 470 ], "score": 1.0, "content": "in different", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 467, 505, 480 ], "spans": [ { "bbox": [ 105, 467, 255, 480 ], "score": 1.0, "content": "branches. It is worth mentioning that", "type": "text" }, { "bbox": [ 256, 468, 271, 480 ], "score": 0.9, "content": "M _ { y }", "type": "inline_equation" }, { "bbox": [ 272, 467, 375, 480 ], "score": 1.0, "content": "here denotes an output of", "type": "text" }, { "bbox": [ 376, 468, 388, 478 ], "score": 0.77, "content": "M", "type": "inline_equation" }, { "bbox": [ 388, 467, 505, 480 ], "score": 1.0, "content": "corresponding to the domain", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 478, 506, 492 ], "spans": [ { "bbox": [ 106, 480, 113, 490 ], "score": 0.68, "content": "y", "type": "inline_equation" }, { "bbox": [ 113, 478, 506, 492 ], "score": 1.0, "content": ". This feature allows our method to separately model the structural appearance (i.e. content) and", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 506, 502 ], "score": 1.0, "content": "its artistic looks (i.e. style), which is essential because applying different styles to the same content", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 500, 504, 514 ], "spans": [ { "bbox": [ 105, 500, 331, 514 ], "score": 1.0, "content": "enriches the diversity of outputs. By randomly sampling", "type": "text" }, { "bbox": [ 331, 502, 338, 511 ], "score": 0.6, "content": "\\mathbf { z }", "type": "inline_equation" }, { "bbox": [ 338, 500, 497, 514 ], "score": 1.0, "content": "from a standard normal distribution and", "type": "text" }, { "bbox": [ 497, 502, 504, 512 ], "score": 0.76, "content": "y", "type": "inline_equation" } ], "index": 22 }, { "bbox": [ 106, 511, 505, 524 ], "spans": [ { "bbox": [ 106, 511, 262, 524 ], "score": 1.0, "content": "from all available foreground domains,", "type": "text" }, { "bbox": [ 263, 512, 275, 521 ], "score": 0.79, "content": "M", "type": "inline_equation" }, { "bbox": [ 275, 511, 505, 524 ], "score": 1.0, "content": "is able to produce diverse style codes and domain specific", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 524, 145, 534 ], "spans": [ { "bbox": [ 106, 524, 145, 534 ], "score": 1.0, "content": "contents.", "type": "text" } ], "index": 24 } ], "index": 20.5, "bbox_fs": [ 105, 444, 506, 534 ] }, { "type": "text", "bbox": [ 109, 539, 503, 573 ], "lines": [ { "bbox": [ 106, 539, 505, 552 ], "spans": [ { "bbox": [ 106, 539, 160, 552 ], "score": 1.0, "content": "The encoder", "type": "text" }, { "bbox": [ 160, 540, 170, 549 ], "score": 0.78, "content": "E", "type": "inline_equation" }, { "bbox": [ 170, 539, 321, 552 ], "score": 1.0, "content": "(Figure 2(c)) extracts the style code", "type": "text" }, { "bbox": [ 322, 540, 370, 551 ], "score": 0.93, "content": "\\mathbf { s } = E _ { y } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 370, 539, 505, 552 ], "score": 1.0, "content": "and the domain specific content", "type": "text" } ], "index": 25 }, { "bbox": [ 107, 550, 505, 563 ], "spans": [ { "bbox": [ 107, 550, 153, 562 ], "score": 0.93, "content": "\\mathbf { c } = E _ { y } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 153, 550, 240, 563 ], "score": 1.0, "content": "from an given image", "type": "text" }, { "bbox": [ 240, 552, 248, 560 ], "score": 0.55, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 248, 550, 505, 563 ], "score": 1.0, "content": ", which reflect the characteristics of reference images instead of", "type": "text" } ], "index": 26 }, { "bbox": [ 107, 561, 209, 574 ], "spans": [ { "bbox": [ 107, 561, 209, 574 ], "score": 1.0, "content": "randomly sampled noise.", "type": "text" } ], "index": 27 } ], "index": 26, "bbox_fs": [ 106, 539, 505, 574 ] }, { "type": "text", "bbox": [ 106, 578, 505, 732 ], "lines": [ { "bbox": [ 106, 578, 505, 591 ], "spans": [ { "bbox": [ 106, 578, 399, 591 ], "score": 1.0, "content": "Foreground/Background (FG/BG) Disentanglement. The generator", "type": "text" }, { "bbox": [ 399, 579, 408, 588 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 409, 578, 505, 591 ], "score": 1.0, "content": "(Figure 2(a)) translates", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 589, 506, 602 ], "spans": [ { "bbox": [ 106, 589, 258, 602 ], "score": 1.0, "content": "an input image x into an output image", "type": "text" }, { "bbox": [ 258, 589, 299, 601 ], "score": 0.93, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 299, 589, 506, 602 ], "score": 1.0, "content": "according to given domain specific style code es and", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 600, 506, 613 ], "spans": [ { "bbox": [ 106, 600, 347, 613 ], "score": 1.0, "content": "content ec, which are provided either by the mapping network", "type": "text" }, { "bbox": [ 348, 601, 360, 610 ], "score": 0.77, "content": "M", "type": "inline_equation" }, { "bbox": [ 360, 600, 506, 613 ], "score": 1.0, "content": "when generating from random noise", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 612, 505, 623 ], "spans": [ { "bbox": [ 106, 612, 232, 623 ], "score": 1.0, "content": "or by the style-content encoder", "type": "text" }, { "bbox": [ 233, 612, 242, 621 ], "score": 0.8, "content": "E", "type": "inline_equation" }, { "bbox": [ 242, 612, 505, 623 ], "score": 1.0, "content": "when transferring an existing content from a reference image. To", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 621, 505, 635 ], "spans": [ { "bbox": [ 105, 621, 505, 635 ], "score": 1.0, "content": "achieve a FG/BG disentanglement, we split the channels of the three-dimensional feature map", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 632, 506, 645 ], "spans": [ { "bbox": [ 105, 632, 128, 645 ], "score": 1.0, "content": "(i.e.,", "type": "text" }, { "bbox": [ 129, 633, 186, 644 ], "score": 0.91, "content": "H \\times W \\times C", "type": "inline_equation" }, { "bbox": [ 186, 632, 278, 645 ], "score": 1.0, "content": ") at the bottle neck of", "type": "text" }, { "bbox": [ 278, 633, 287, 643 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 288, 632, 506, 645 ], "score": 1.0, "content": "into two parts. The model is then forced to encode", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 643, 505, 657 ], "spans": [ { "bbox": [ 105, 643, 505, 657 ], "score": 1.0, "content": "the background into the first channels and the domain specific content cˆ into the latter channels by", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 654, 506, 668 ], "spans": [ { "bbox": [ 105, 654, 506, 668 ], "score": 1.0, "content": "classification losses as discussed in Section 3.2. cˆ can then be replaced with content ec from the target", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 665, 506, 679 ], "spans": [ { "bbox": [ 105, 665, 506, 679 ], "score": 1.0, "content": "domain. The adaptive instance normalization (AdaIN) (Huang & Belongie, 2017) is then used to", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 677, 505, 690 ], "spans": [ { "bbox": [ 106, 677, 374, 690 ], "score": 1.0, "content": "inject es into ec during the decoding process while the background", "type": "text" }, { "bbox": [ 375, 677, 412, 689 ], "score": 0.92, "content": "B G _ { G } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 412, 677, 505, 690 ], "score": 1.0, "content": "is decoded separately.", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 187, 700 ], "score": 1.0, "content": "StarGAN v2 learns", "type": "text" }, { "bbox": [ 187, 688, 204, 698 ], "score": 0.31, "content": "\\mathbf { F G }", "type": "inline_equation" }, { "bbox": [ 204, 687, 505, 700 ], "score": 1.0, "content": "and BG together which leads to a conditional relationship between both.", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Our disentanglement and separate encoding break this conditioning and therefore enable our method", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 709, 504, 722 ], "spans": [ { "bbox": [ 105, 710, 466, 722 ], "score": 1.0, "content": "to freely combine FG and BG as well as learn the full variation of FG content. Finally,", "type": "text" }, { "bbox": [ 466, 709, 504, 722 ], "score": 0.92, "content": "B G _ { G } ( \\mathbf { x } )", "type": "inline_equation" } ], "index": 40 }, { "bbox": [ 106, 721, 354, 733 ], "spans": [ { "bbox": [ 106, 721, 354, 733 ], "score": 1.0, "content": "and ec are concatenated together and then fused before output.", "type": "text" } ], "index": 41 } ], "index": 34.5, "bbox_fs": [ 105, 578, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 171 ], "lines": [ { "bbox": [ 106, 82, 504, 95 ], "spans": [ { "bbox": [ 106, 82, 397, 95 ], "score": 1.0, "content": "Multi-task discriminator with auxiliary classifiers. The discriminator", "type": "text" }, { "bbox": [ 397, 83, 407, 93 ], "score": 0.75, "content": "D", "type": "inline_equation" }, { "bbox": [ 407, 82, 504, 95 ], "score": 1.0, "content": "(Figure 2(d)) is a multi-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "task discriminator with two auxiliary classifiers: a foreground domain classifier and a background", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "domain classifier. This feature strengthens the disentanglement of FG and BG by first ensuring the", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 156, 128 ], "score": 1.0, "content": "input image", "type": "text" }, { "bbox": [ 156, 117, 164, 126 ], "score": 0.27, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 165, 115, 505, 128 ], "score": 1.0, "content": "contains a domain specific content that can be recognized by the foreground domain", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 127, 506, 139 ], "spans": [ { "bbox": [ 106, 127, 361, 139 ], "score": 1.0, "content": "classifier independent of the background. Later, each branch", "type": "text" }, { "bbox": [ 361, 127, 376, 138 ], "score": 0.89, "content": "D _ { y }", "type": "inline_equation" }, { "bbox": [ 376, 127, 506, 139 ], "score": 1.0, "content": "in the multi-task discriminator", "type": "text" } ], "index": 4 }, { "bbox": [ 107, 136, 505, 151 ], "spans": [ { "bbox": [ 107, 137, 116, 147 ], "score": 0.8, "content": "D", "type": "inline_equation" }, { "bbox": [ 117, 136, 262, 151 ], "score": 1.0, "content": "is trained to determine if an image", "type": "text" }, { "bbox": [ 262, 139, 270, 147 ], "score": 0.43, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 270, 136, 505, 151 ], "score": 1.0, "content": "is a real image of its foreground domain or a fake image", "type": "text" } ], "index": 5 }, { "bbox": [ 107, 147, 506, 161 ], "spans": [ { "bbox": [ 107, 148, 147, 160 ], "score": 0.91, "content": "G ( \\mathbf { x } , \\mathbf { s } , \\mathbf { c } )", "type": "inline_equation" }, { "bbox": [ 148, 147, 204, 161 ], "score": 1.0, "content": "generated by", "type": "text" }, { "bbox": [ 204, 149, 213, 158 ], "score": 0.72, "content": "G", "type": "inline_equation" }, { "bbox": [ 214, 147, 358, 161 ], "score": 1.0, "content": ". Apart from that, one extra branch", "type": "text" }, { "bbox": [ 358, 149, 383, 159 ], "score": 0.91, "content": "B G _ { \\mathrm { c l s } }", "type": "inline_equation" }, { "bbox": [ 384, 147, 506, 161 ], "score": 1.0, "content": "is attached to decide whether", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 159, 371, 172 ], "spans": [ { "bbox": [ 106, 159, 371, 172 ], "score": 1.0, "content": "the background information of the input images is well preserved.", "type": "text" } ], "index": 7 } ], "index": 3.5 }, { "type": "text", "bbox": [ 106, 176, 505, 254 ], "lines": [ { "bbox": [ 106, 176, 505, 188 ], "spans": [ { "bbox": [ 106, 176, 505, 188 ], "score": 1.0, "content": "Content Transfer. Mokady et al. (2020) introduced a concept that a model should be able to identity", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 186, 505, 199 ], "spans": [ { "bbox": [ 106, 186, 505, 199 ], "score": 1.0, "content": "the difference between two domains when one of the domains contains a feature that the other does", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 198, 505, 210 ], "spans": [ { "bbox": [ 105, 198, 418, 210 ], "score": 1.0, "content": "not have. We refer to this concept as ‘anchor’ and extend to multiple domains", "type": "text" }, { "bbox": [ 418, 199, 438, 209 ], "score": 0.83, "content": "( > 2 )", "type": "inline_equation" }, { "bbox": [ 438, 198, 505, 210 ], "score": 1.0, "content": "by the FG/BG", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 210, 505, 221 ], "spans": [ { "bbox": [ 106, 210, 453, 221 ], "score": 1.0, "content": "disentanglement, the multi-task discriminator and the foreground content classifier in", "type": "text" }, { "bbox": [ 453, 210, 463, 219 ], "score": 0.7, "content": "D", "type": "inline_equation" }, { "bbox": [ 463, 210, 505, 221 ], "score": 1.0, "content": ". We treat", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 221, 505, 232 ], "spans": [ { "bbox": [ 106, 221, 505, 232 ], "score": 1.0, "content": "domain Normal as the anchor domain i.e. set the domain specific content to zero, because a normal", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 232, 505, 243 ], "spans": [ { "bbox": [ 106, 232, 505, 243 ], "score": 1.0, "content": "image has no domain specific content in our definition. As a result, we can now transfer contents", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 241, 374, 254 ], "spans": [ { "bbox": [ 106, 241, 374, 254 ], "score": 1.0, "content": "between all combination of FG and BG domains (see Figure 10).", "type": "text" } ], "index": 14 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 258, 505, 314 ], "lines": [ { "bbox": [ 106, 258, 505, 271 ], "spans": [ { "bbox": [ 106, 258, 505, 271 ], "score": 1.0, "content": "Compared to StarGAN v2, our method not only models style codes and contents separately but also", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 268, 505, 282 ], "spans": [ { "bbox": [ 105, 268, 505, 282 ], "score": 1.0, "content": "disentangles the foreground and background of an image in a weakly-supervised manner. These", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 281, 505, 294 ], "spans": [ { "bbox": [ 105, 281, 505, 294 ], "score": 1.0, "content": "features allow explicit control over output images by combining desired style codes and contents", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 290, 505, 306 ], "spans": [ { "bbox": [ 105, 290, 505, 306 ], "score": 1.0, "content": "from one of the subnetworks with the input images. Therefore, it leads to higher variance regarding", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 303, 485, 316 ], "spans": [ { "bbox": [ 105, 303, 485, 316 ], "score": 1.0, "content": "the location, structural pattern and artistic style of defects in the synthetic images of DT-GAN.", "type": "text" } ], "index": 19 } ], "index": 17 }, { "type": "title", "bbox": [ 108, 327, 232, 339 ], "lines": [ { "bbox": [ 106, 327, 232, 340 ], "spans": [ { "bbox": [ 106, 327, 232, 340 ], "score": 1.0, "content": "3.2 TRAINING OBJECTIVES", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 348, 504, 371 ], "lines": [ { "bbox": [ 105, 346, 506, 361 ], "spans": [ { "bbox": [ 105, 346, 172, 361 ], "score": 1.0, "content": "Given an image", "type": "text" }, { "bbox": [ 172, 349, 202, 359 ], "score": 0.89, "content": "\\mathbf { x } \\in \\mathcal { X }", "type": "inline_equation" }, { "bbox": [ 202, 346, 332, 361 ], "score": 1.0, "content": ", its original foreground domain", "type": "text" }, { "bbox": [ 332, 349, 360, 360 ], "score": 0.91, "content": "y \\in \\mathcal { V }", "type": "inline_equation" }, { "bbox": [ 361, 346, 473, 361 ], "score": 1.0, "content": "and its background domain", "type": "text" }, { "bbox": [ 474, 348, 501, 360 ], "score": 0.9, "content": "p \\in \\mathcal { P }", "type": "inline_equation" }, { "bbox": [ 501, 346, 506, 361 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 360, 335, 371 ], "spans": [ { "bbox": [ 106, 360, 335, 371 ], "score": 1.0, "content": "the following objectives are used to train our framework.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "text", "bbox": [ 106, 376, 505, 453 ], "lines": [ { "bbox": [ 106, 376, 504, 389 ], "spans": [ { "bbox": [ 106, 376, 323, 389 ], "score": 1.0, "content": "Adversarial loss. In the training phase, a noise vector", "type": "text" }, { "bbox": [ 323, 377, 349, 387 ], "score": 0.9, "content": "\\mathbf { z } \\in { \\mathcal { Z } }", "type": "inline_equation" }, { "bbox": [ 350, 376, 478, 389 ], "score": 1.0, "content": "and a target foreground domain", "type": "text" }, { "bbox": [ 478, 376, 504, 388 ], "score": 0.9, "content": "\\widetilde y \\in \\mathcal { V }", "type": "inline_equation" } ], "index": 23 }, { "bbox": [ 105, 387, 505, 400 ], "spans": [ { "bbox": [ 105, 387, 306, 400 ], "score": 1.0, "content": "are sampled randomly. Both of them are fed to", "type": "text" }, { "bbox": [ 307, 388, 318, 397 ], "score": 0.82, "content": "M", "type": "inline_equation" }, { "bbox": [ 319, 387, 445, 400 ], "score": 1.0, "content": ", producing a target style code", "type": "text" }, { "bbox": [ 446, 388, 453, 397 ], "score": 0.66, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 453, 387, 505, 400 ], "score": 1.0, "content": "and a target", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 398, 505, 411 ], "spans": [ { "bbox": [ 105, 398, 192, 411 ], "score": 1.0, "content": "content ec as follows:", "type": "text" }, { "bbox": [ 193, 398, 249, 410 ], "score": 0.93, "content": "\\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } = M _ { \\widetilde { y } } ( \\mathbf { z } )", "type": "inline_equation" }, { "bbox": [ 249, 398, 400, 411 ], "score": 1.0, "content": ". Goal of the training is to ensure that", "type": "text" }, { "bbox": [ 401, 398, 407, 408 ], "score": 0.48, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 407, 398, 505, 411 ], "score": 1.0, "content": "and ec are sampled from", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 409, 505, 421 ], "spans": [ { "bbox": [ 106, 410, 359, 421 ], "score": 1.0, "content": "the distribution over styles and contents of the target domain", "type": "text" }, { "bbox": [ 359, 409, 366, 420 ], "score": 0.84, "content": "\\widetilde { y }", "type": "inline_equation" }, { "bbox": [ 366, 410, 432, 421 ], "score": 1.0, "content": ". The generator", "type": "text" }, { "bbox": [ 433, 410, 442, 419 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 442, 410, 505, 421 ], "score": 1.0, "content": "then combines", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 420, 504, 432 ], "spans": [ { "bbox": [ 105, 420, 145, 432 ], "score": 1.0, "content": "an image", "type": "text" }, { "bbox": [ 145, 421, 154, 430 ], "score": 0.57, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 154, 420, 174, 432 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 174, 420, 181, 430 ], "score": 0.35, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 182, 420, 199, 432 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 199, 420, 206, 430 ], "score": 0.27, "content": "\\widetilde { \\mathbf c }", "type": "inline_equation" }, { "bbox": [ 206, 420, 365, 432 ], "score": 1.0, "content": "and learns to generate an output image", "type": "text" }, { "bbox": [ 365, 420, 406, 432 ], "score": 0.93, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 406, 420, 504, 432 ], "score": 1.0, "content": "that is indistinguishable", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 431, 505, 443 ], "spans": [ { "bbox": [ 106, 431, 262, 443 ], "score": 1.0, "content": "from real images in the target domain", "type": "text" }, { "bbox": [ 263, 431, 269, 443 ], "score": 0.82, "content": "\\widetilde { y }", "type": "inline_equation" }, { "bbox": [ 270, 431, 505, 443 ], "score": 1.0, "content": ". We encourage this behavior by using an adversarial loss", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 442, 223, 454 ], "spans": [ { "bbox": [ 105, 442, 223, 454 ], "score": 1.0, "content": "same as in Choi et al. (2020)", "type": "text" } ], "index": 29 } ], "index": 26 }, { "type": "interline_equation", "bbox": [ 178, 459, 432, 474 ], "lines": [ { "bbox": [ 178, 459, 432, 474 ], "spans": [ { "bbox": [ 178, 459, 432, 474 ], "score": 0.9, "content": "\\mathcal { L } _ { \\mathrm { a d v } } = \\mathbb { E } _ { { \\mathbf { x } } , y } \\big [ \\log D _ { y } ( { \\mathbf { x } } ) \\big ] + \\mathbb { E } _ { { \\mathbf { x } } , \\widetilde { y } , { \\mathbf { z } } } [ \\log \\left( 1 - D _ { \\widetilde { y } } ( G ( { \\mathbf { x } } , \\widetilde { { \\mathbf { s } } } , \\widetilde { { \\mathbf { c } } } ) ) \\right) ] ,", "type": "interline_equation", "image_path": "cd793fc06d67b12ed9b5a25f7cc4effb3e16a9336e8b3693d7bd6c236854a0d3.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 178, 459, 432, 474 ], "spans": [], "index": 30 } ] }, { "type": "text", "bbox": [ 106, 480, 504, 504 ], "lines": [ { "bbox": [ 106, 479, 506, 493 ], "spans": [ { "bbox": [ 106, 479, 134, 493 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 481, 149, 492 ], "score": 0.9, "content": "D _ { y }", "type": "inline_equation" }, { "bbox": [ 149, 479, 168, 493 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 169, 481, 183, 493 ], "score": 0.9, "content": "D _ { \\widetilde { y } }", "type": "inline_equation" }, { "bbox": [ 183, 479, 296, 493 ], "score": 1.0, "content": "are the output branches of", "type": "text" }, { "bbox": [ 296, 481, 306, 490 ], "score": 0.84, "content": "D", "type": "inline_equation" }, { "bbox": [ 306, 479, 464, 493 ], "score": 1.0, "content": "that correspond to the source domain", "type": "text" }, { "bbox": [ 464, 483, 470, 492 ], "score": 0.79, "content": "y", "type": "inline_equation" }, { "bbox": [ 471, 479, 506, 493 ], "score": 1.0, "content": "and the", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 492, 226, 503 ], "spans": [ { "bbox": [ 106, 492, 164, 503 ], "score": 1.0, "content": "target domain", "type": "text" }, { "bbox": [ 164, 492, 170, 503 ], "score": 0.82, "content": "\\widetilde { y }", "type": "inline_equation" }, { "bbox": [ 171, 492, 226, 503 ], "score": 1.0, "content": ", respectively.", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "text", "bbox": [ 106, 507, 505, 542 ], "lines": [ { "bbox": [ 105, 507, 505, 521 ], "spans": [ { "bbox": [ 105, 507, 455, 521 ], "score": 1.0, "content": "Style-content reconstruction loss. Similar to StarGAN v2, to enforce the generator", "type": "text" }, { "bbox": [ 456, 509, 465, 518 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 465, 507, 505, 521 ], "score": 1.0, "content": "takes the", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 519, 505, 532 ], "spans": [ { "bbox": [ 105, 519, 149, 532 ], "score": 1.0, "content": "style code", "type": "text" }, { "bbox": [ 149, 519, 156, 529 ], "score": 0.64, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 156, 519, 505, 532 ], "score": 1.0, "content": "and the domain specific content ec into consideration during the generation process, we", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 531, 278, 543 ], "spans": [ { "bbox": [ 106, 531, 278, 543 ], "score": 1.0, "content": "employ a style-content reconstruction loss", "type": "text" } ], "index": 35 } ], "index": 34 }, { "type": "interline_equation", "bbox": [ 148, 547, 463, 563 ], "lines": [ { "bbox": [ 148, 547, 463, 563 ], "spans": [ { "bbox": [ 148, 547, 463, 563 ], "score": 0.89, "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { s t y . c o n } } = \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } } \\big [ \\| \\widetilde { \\mathbf { s } } - S _ { E } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) ) \\| _ { 1 } \\big ] + \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } } \\big [ \\| \\widetilde { \\mathbf { c } } - C _ { E } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) ) \\| _ { 1 } \\big ] . } \\end{array}", "type": "interline_equation", "image_path": "ae70d605b176b542cef1b090f291d2c692ecbbaaceb1c0f0b119e790a5f72661.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 148, 547, 463, 563 ], "spans": [], "index": 36 } ] }, { "type": "text", "bbox": [ 106, 568, 505, 603 ], "lines": [ { "bbox": [ 106, 568, 505, 581 ], "spans": [ { "bbox": [ 106, 569, 291, 581 ], "score": 1.0, "content": "This objective urges the style-content encoder", "type": "text" }, { "bbox": [ 291, 569, 300, 579 ], "score": 0.84, "content": "E", "type": "inline_equation" }, { "bbox": [ 300, 569, 342, 581 ], "score": 1.0, "content": "to recover", "type": "text" }, { "bbox": [ 343, 568, 350, 579 ], "score": 0.47, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 350, 569, 396, 581 ], "score": 1.0, "content": "and c from", "type": "text" }, { "bbox": [ 396, 568, 437, 581 ], "score": 0.91, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 437, 569, 505, 581 ], "score": 1.0, "content": ". Here, the style-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 579, 506, 593 ], "spans": [ { "bbox": [ 105, 579, 173, 593 ], "score": 1.0, "content": "content encoder", "type": "text" }, { "bbox": [ 173, 580, 182, 590 ], "score": 0.83, "content": "E", "type": "inline_equation" }, { "bbox": [ 183, 579, 506, 593 ], "score": 1.0, "content": "learns a mapping from an image to its style and content domains, which allows", "type": "text" } ], "index": 38 }, { "bbox": [ 107, 590, 426, 603 ], "spans": [ { "bbox": [ 107, 592, 115, 601 ], "score": 0.8, "content": "G", "type": "inline_equation" }, { "bbox": [ 116, 590, 426, 603 ], "score": 1.0, "content": "to synthesize an image with given s and c from reference images at test time.", "type": "text" } ], "index": 39 } ], "index": 38 }, { "type": "text", "bbox": [ 107, 607, 504, 652 ], "lines": [ { "bbox": [ 105, 606, 506, 620 ], "spans": [ { "bbox": [ 105, 606, 427, 620 ], "score": 1.0, "content": "Diversity loss. In order to further boost the diversity of output images from", "type": "text" }, { "bbox": [ 428, 608, 436, 618 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 437, 606, 506, 620 ], "score": 1.0, "content": ", we introduce a", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 619, 505, 631 ], "spans": [ { "bbox": [ 105, 619, 411, 631 ], "score": 1.0, "content": "loss that encourages diversity as follows: for a pair of random latent codes", "type": "text" }, { "bbox": [ 411, 620, 422, 630 ], "score": 0.85, "content": "\\mathbf { z } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 423, 619, 441, 631 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 441, 620, 452, 630 ], "score": 0.87, "content": "\\mathbf { z } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 452, 619, 505, 631 ], "score": 1.0, "content": "we compute", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 630, 504, 642 ], "spans": [ { "bbox": [ 106, 630, 174, 642 ], "score": 0.93, "content": "\\widetilde { \\mathbf { s } } _ { i } , \\widetilde { \\mathbf { c } } _ { i } = M _ { \\widetilde { y } } ( \\mathbf { z } _ { i } )", "type": "inline_equation" }, { "bbox": [ 174, 630, 191, 642 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 191, 630, 234, 642 ], "score": 0.93, "content": "i \\in \\{ 1 , 2 \\}", "type": "inline_equation" }, { "bbox": [ 234, 630, 435, 642 ], "score": 1.0, "content": "and enforce a different outcome of the generator", "type": "text" }, { "bbox": [ 435, 630, 444, 640 ], "score": 0.84, "content": "G", "type": "inline_equation" }, { "bbox": [ 444, 630, 504, 642 ], "score": 1.0, "content": "for differently", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 640, 252, 653 ], "spans": [ { "bbox": [ 105, 640, 252, 653 ], "score": 1.0, "content": "mixed style and content input pairs:", "type": "text" } ], "index": 43 } ], "index": 41.5 }, { "type": "interline_equation", "bbox": [ 182, 656, 428, 706 ], "lines": [ { "bbox": [ 182, 656, 428, 706 ], "spans": [ { "bbox": [ 182, 656, 428, 706 ], "score": 0.93, "content": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { d s } } = \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } } \\left[ \\| G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 1 } , \\widetilde { \\mathbf { c } } _ { 2 } ) - G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 2 } , \\widetilde { \\mathbf { c } } _ { 1 } ) \\| _ { 1 } \\right] } \\\\ & { \\quad \\quad + \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } } \\left[ \\| G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 1 } , \\widetilde { \\mathbf { c } } _ { 1 } ) - G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 2 } , \\widetilde { \\mathbf { c } } _ { 2 } ) \\| _ { 1 } \\right] } \\\\ & { \\quad \\quad + \\sum _ { m , n , o } \\left[ \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } } \\left[ \\| G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { m } , \\widetilde { \\mathbf { c } } _ { n } ) - G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { o } , \\widetilde { \\mathbf { c } } _ { o } ) \\| _ { 1 } \\right] \\right] , } \\end{array}", "type": "interline_equation", "image_path": "af2defdd3af95b358356d3f45b7f10b9e2ba3e2f9851f7366b219e39a10c234c.jpg" } ] } ], "index": 45, "virtual_lines": [ { "bbox": [ 182, 656, 428, 672.6666666666666 ], "spans": [], "index": 44 }, { "bbox": [ 182, 672.6666666666666, 428, 689.3333333333333 ], "spans": [], "index": 45 }, { "bbox": [ 182, 689.3333333333333, 428, 705.9999999999999 ], "spans": [], "index": 46 } ] }, { "type": "text", "bbox": [ 108, 709, 503, 732 ], "lines": [ { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 134, 722 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 709, 228, 722 ], "score": 0.92, "content": "m , n \\in \\{ 1 , 2 | m \\neq n \\}", "type": "inline_equation" }, { "bbox": [ 229, 709, 249, 722 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 249, 710, 294, 722 ], "score": 0.92, "content": "o \\in \\{ 1 , 2 \\}", "type": "inline_equation" }, { "bbox": [ 295, 709, 444, 722 ], "score": 1.0, "content": ". Driven by this term, the generator", "type": "text" }, { "bbox": [ 445, 710, 454, 720 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 454, 709, 505, 722 ], "score": 1.0, "content": "is forced to", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 719, 505, 734 ], "spans": [ { "bbox": [ 106, 719, 505, 734 ], "score": 1.0, "content": "discover meaningful style features and contents that eventually lead to diversity in generated images.", "type": "text" } ], "index": 48 } ], "index": 47.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "4", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 171 ], "lines": [ { "bbox": [ 106, 82, 504, 95 ], "spans": [ { "bbox": [ 106, 82, 397, 95 ], "score": 1.0, "content": "Multi-task discriminator with auxiliary classifiers. The discriminator", "type": "text" }, { "bbox": [ 397, 83, 407, 93 ], "score": 0.75, "content": "D", "type": "inline_equation" }, { "bbox": [ 407, 82, 504, 95 ], "score": 1.0, "content": "(Figure 2(d)) is a multi-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "task discriminator with two auxiliary classifiers: a foreground domain classifier and a background", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "domain classifier. This feature strengthens the disentanglement of FG and BG by first ensuring the", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 156, 128 ], "score": 1.0, "content": "input image", "type": "text" }, { "bbox": [ 156, 117, 164, 126 ], "score": 0.27, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 165, 115, 505, 128 ], "score": 1.0, "content": "contains a domain specific content that can be recognized by the foreground domain", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 127, 506, 139 ], "spans": [ { "bbox": [ 106, 127, 361, 139 ], "score": 1.0, "content": "classifier independent of the background. Later, each branch", "type": "text" }, { "bbox": [ 361, 127, 376, 138 ], "score": 0.89, "content": "D _ { y }", "type": "inline_equation" }, { "bbox": [ 376, 127, 506, 139 ], "score": 1.0, "content": "in the multi-task discriminator", "type": "text" } ], "index": 4 }, { "bbox": [ 107, 136, 505, 151 ], "spans": [ { "bbox": [ 107, 137, 116, 147 ], "score": 0.8, "content": "D", "type": "inline_equation" }, { "bbox": [ 117, 136, 262, 151 ], "score": 1.0, "content": "is trained to determine if an image", "type": "text" }, { "bbox": [ 262, 139, 270, 147 ], "score": 0.43, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 270, 136, 505, 151 ], "score": 1.0, "content": "is a real image of its foreground domain or a fake image", "type": "text" } ], "index": 5 }, { "bbox": [ 107, 147, 506, 161 ], "spans": [ { "bbox": [ 107, 148, 147, 160 ], "score": 0.91, "content": "G ( \\mathbf { x } , \\mathbf { s } , \\mathbf { c } )", "type": "inline_equation" }, { "bbox": [ 148, 147, 204, 161 ], "score": 1.0, "content": "generated by", "type": "text" }, { "bbox": [ 204, 149, 213, 158 ], "score": 0.72, "content": "G", "type": "inline_equation" }, { "bbox": [ 214, 147, 358, 161 ], "score": 1.0, "content": ". Apart from that, one extra branch", "type": "text" }, { "bbox": [ 358, 149, 383, 159 ], "score": 0.91, "content": "B G _ { \\mathrm { c l s } }", "type": "inline_equation" }, { "bbox": [ 384, 147, 506, 161 ], "score": 1.0, "content": "is attached to decide whether", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 159, 371, 172 ], "spans": [ { "bbox": [ 106, 159, 371, 172 ], "score": 1.0, "content": "the background information of the input images is well preserved.", "type": "text" } ], "index": 7 } ], "index": 3.5, "bbox_fs": [ 105, 82, 506, 172 ] }, { "type": "text", "bbox": [ 106, 176, 505, 254 ], "lines": [ { "bbox": [ 106, 176, 505, 188 ], "spans": [ { "bbox": [ 106, 176, 505, 188 ], "score": 1.0, "content": "Content Transfer. Mokady et al. (2020) introduced a concept that a model should be able to identity", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 186, 505, 199 ], "spans": [ { "bbox": [ 106, 186, 505, 199 ], "score": 1.0, "content": "the difference between two domains when one of the domains contains a feature that the other does", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 198, 505, 210 ], "spans": [ { "bbox": [ 105, 198, 418, 210 ], "score": 1.0, "content": "not have. We refer to this concept as ‘anchor’ and extend to multiple domains", "type": "text" }, { "bbox": [ 418, 199, 438, 209 ], "score": 0.83, "content": "( > 2 )", "type": "inline_equation" }, { "bbox": [ 438, 198, 505, 210 ], "score": 1.0, "content": "by the FG/BG", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 210, 505, 221 ], "spans": [ { "bbox": [ 106, 210, 453, 221 ], "score": 1.0, "content": "disentanglement, the multi-task discriminator and the foreground content classifier in", "type": "text" }, { "bbox": [ 453, 210, 463, 219 ], "score": 0.7, "content": "D", "type": "inline_equation" }, { "bbox": [ 463, 210, 505, 221 ], "score": 1.0, "content": ". We treat", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 221, 505, 232 ], "spans": [ { "bbox": [ 106, 221, 505, 232 ], "score": 1.0, "content": "domain Normal as the anchor domain i.e. set the domain specific content to zero, because a normal", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 232, 505, 243 ], "spans": [ { "bbox": [ 106, 232, 505, 243 ], "score": 1.0, "content": "image has no domain specific content in our definition. As a result, we can now transfer contents", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 241, 374, 254 ], "spans": [ { "bbox": [ 106, 241, 374, 254 ], "score": 1.0, "content": "between all combination of FG and BG domains (see Figure 10).", "type": "text" } ], "index": 14 } ], "index": 11, "bbox_fs": [ 105, 176, 505, 254 ] }, { "type": "text", "bbox": [ 107, 258, 505, 314 ], "lines": [ { "bbox": [ 106, 258, 505, 271 ], "spans": [ { "bbox": [ 106, 258, 505, 271 ], "score": 1.0, "content": "Compared to StarGAN v2, our method not only models style codes and contents separately but also", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 268, 505, 282 ], "spans": [ { "bbox": [ 105, 268, 505, 282 ], "score": 1.0, "content": "disentangles the foreground and background of an image in a weakly-supervised manner. These", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 281, 505, 294 ], "spans": [ { "bbox": [ 105, 281, 505, 294 ], "score": 1.0, "content": "features allow explicit control over output images by combining desired style codes and contents", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 290, 505, 306 ], "spans": [ { "bbox": [ 105, 290, 505, 306 ], "score": 1.0, "content": "from one of the subnetworks with the input images. Therefore, it leads to higher variance regarding", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 303, 485, 316 ], "spans": [ { "bbox": [ 105, 303, 485, 316 ], "score": 1.0, "content": "the location, structural pattern and artistic style of defects in the synthetic images of DT-GAN.", "type": "text" } ], "index": 19 } ], "index": 17, "bbox_fs": [ 105, 258, 505, 316 ] }, { "type": "title", "bbox": [ 108, 327, 232, 339 ], "lines": [ { "bbox": [ 106, 327, 232, 340 ], "spans": [ { "bbox": [ 106, 327, 232, 340 ], "score": 1.0, "content": "3.2 TRAINING OBJECTIVES", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 348, 504, 371 ], "lines": [ { "bbox": [ 105, 346, 506, 361 ], "spans": [ { "bbox": [ 105, 346, 172, 361 ], "score": 1.0, "content": "Given an image", "type": "text" }, { "bbox": [ 172, 349, 202, 359 ], "score": 0.89, "content": "\\mathbf { x } \\in \\mathcal { X }", "type": "inline_equation" }, { "bbox": [ 202, 346, 332, 361 ], "score": 1.0, "content": ", its original foreground domain", "type": "text" }, { "bbox": [ 332, 349, 360, 360 ], "score": 0.91, "content": "y \\in \\mathcal { V }", "type": "inline_equation" }, { "bbox": [ 361, 346, 473, 361 ], "score": 1.0, "content": "and its background domain", "type": "text" }, { "bbox": [ 474, 348, 501, 360 ], "score": 0.9, "content": "p \\in \\mathcal { P }", "type": "inline_equation" }, { "bbox": [ 501, 346, 506, 361 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 360, 335, 371 ], "spans": [ { "bbox": [ 106, 360, 335, 371 ], "score": 1.0, "content": "the following objectives are used to train our framework.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 105, 346, 506, 371 ] }, { "type": "text", "bbox": [ 106, 376, 505, 453 ], "lines": [ { "bbox": [ 106, 376, 504, 389 ], "spans": [ { "bbox": [ 106, 376, 323, 389 ], "score": 1.0, "content": "Adversarial loss. In the training phase, a noise vector", "type": "text" }, { "bbox": [ 323, 377, 349, 387 ], "score": 0.9, "content": "\\mathbf { z } \\in { \\mathcal { Z } }", "type": "inline_equation" }, { "bbox": [ 350, 376, 478, 389 ], "score": 1.0, "content": "and a target foreground domain", "type": "text" }, { "bbox": [ 478, 376, 504, 388 ], "score": 0.9, "content": "\\widetilde y \\in \\mathcal { V }", "type": "inline_equation" } ], "index": 23 }, { "bbox": [ 105, 387, 505, 400 ], "spans": [ { "bbox": [ 105, 387, 306, 400 ], "score": 1.0, "content": "are sampled randomly. Both of them are fed to", "type": "text" }, { "bbox": [ 307, 388, 318, 397 ], "score": 0.82, "content": "M", "type": "inline_equation" }, { "bbox": [ 319, 387, 445, 400 ], "score": 1.0, "content": ", producing a target style code", "type": "text" }, { "bbox": [ 446, 388, 453, 397 ], "score": 0.66, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 453, 387, 505, 400 ], "score": 1.0, "content": "and a target", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 398, 505, 411 ], "spans": [ { "bbox": [ 105, 398, 192, 411 ], "score": 1.0, "content": "content ec as follows:", "type": "text" }, { "bbox": [ 193, 398, 249, 410 ], "score": 0.93, "content": "\\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } = M _ { \\widetilde { y } } ( \\mathbf { z } )", "type": "inline_equation" }, { "bbox": [ 249, 398, 400, 411 ], "score": 1.0, "content": ". Goal of the training is to ensure that", "type": "text" }, { "bbox": [ 401, 398, 407, 408 ], "score": 0.48, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 407, 398, 505, 411 ], "score": 1.0, "content": "and ec are sampled from", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 409, 505, 421 ], "spans": [ { "bbox": [ 106, 410, 359, 421 ], "score": 1.0, "content": "the distribution over styles and contents of the target domain", "type": "text" }, { "bbox": [ 359, 409, 366, 420 ], "score": 0.84, "content": "\\widetilde { y }", "type": "inline_equation" }, { "bbox": [ 366, 410, 432, 421 ], "score": 1.0, "content": ". The generator", "type": "text" }, { "bbox": [ 433, 410, 442, 419 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 442, 410, 505, 421 ], "score": 1.0, "content": "then combines", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 420, 504, 432 ], "spans": [ { "bbox": [ 105, 420, 145, 432 ], "score": 1.0, "content": "an image", "type": "text" }, { "bbox": [ 145, 421, 154, 430 ], "score": 0.57, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 154, 420, 174, 432 ], "score": 1.0, "content": "with", "type": "text" }, { "bbox": [ 174, 420, 181, 430 ], "score": 0.35, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 182, 420, 199, 432 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 199, 420, 206, 430 ], "score": 0.27, "content": "\\widetilde { \\mathbf c }", "type": "inline_equation" }, { "bbox": [ 206, 420, 365, 432 ], "score": 1.0, "content": "and learns to generate an output image", "type": "text" }, { "bbox": [ 365, 420, 406, 432 ], "score": 0.93, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 406, 420, 504, 432 ], "score": 1.0, "content": "that is indistinguishable", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 431, 505, 443 ], "spans": [ { "bbox": [ 106, 431, 262, 443 ], "score": 1.0, "content": "from real images in the target domain", "type": "text" }, { "bbox": [ 263, 431, 269, 443 ], "score": 0.82, "content": "\\widetilde { y }", "type": "inline_equation" }, { "bbox": [ 270, 431, 505, 443 ], "score": 1.0, "content": ". We encourage this behavior by using an adversarial loss", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 442, 223, 454 ], "spans": [ { "bbox": [ 105, 442, 223, 454 ], "score": 1.0, "content": "same as in Choi et al. (2020)", "type": "text" } ], "index": 29 } ], "index": 26, "bbox_fs": [ 105, 376, 505, 454 ] }, { "type": "interline_equation", "bbox": [ 178, 459, 432, 474 ], "lines": [ { "bbox": [ 178, 459, 432, 474 ], "spans": [ { "bbox": [ 178, 459, 432, 474 ], "score": 0.9, "content": "\\mathcal { L } _ { \\mathrm { a d v } } = \\mathbb { E } _ { { \\mathbf { x } } , y } \\big [ \\log D _ { y } ( { \\mathbf { x } } ) \\big ] + \\mathbb { E } _ { { \\mathbf { x } } , \\widetilde { y } , { \\mathbf { z } } } [ \\log \\left( 1 - D _ { \\widetilde { y } } ( G ( { \\mathbf { x } } , \\widetilde { { \\mathbf { s } } } , \\widetilde { { \\mathbf { c } } } ) ) \\right) ] ,", "type": "interline_equation", "image_path": "cd793fc06d67b12ed9b5a25f7cc4effb3e16a9336e8b3693d7bd6c236854a0d3.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 178, 459, 432, 474 ], "spans": [], "index": 30 } ] }, { "type": "text", "bbox": [ 106, 480, 504, 504 ], "lines": [ { "bbox": [ 106, 479, 506, 493 ], "spans": [ { "bbox": [ 106, 479, 134, 493 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 481, 149, 492 ], "score": 0.9, "content": "D _ { y }", "type": "inline_equation" }, { "bbox": [ 149, 479, 168, 493 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 169, 481, 183, 493 ], "score": 0.9, "content": "D _ { \\widetilde { y } }", "type": "inline_equation" }, { "bbox": [ 183, 479, 296, 493 ], "score": 1.0, "content": "are the output branches of", "type": "text" }, { "bbox": [ 296, 481, 306, 490 ], "score": 0.84, "content": "D", "type": "inline_equation" }, { "bbox": [ 306, 479, 464, 493 ], "score": 1.0, "content": "that correspond to the source domain", "type": "text" }, { "bbox": [ 464, 483, 470, 492 ], "score": 0.79, "content": "y", "type": "inline_equation" }, { "bbox": [ 471, 479, 506, 493 ], "score": 1.0, "content": "and the", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 492, 226, 503 ], "spans": [ { "bbox": [ 106, 492, 164, 503 ], "score": 1.0, "content": "target domain", "type": "text" }, { "bbox": [ 164, 492, 170, 503 ], "score": 0.82, "content": "\\widetilde { y }", "type": "inline_equation" }, { "bbox": [ 171, 492, 226, 503 ], "score": 1.0, "content": ", respectively.", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 106, 479, 506, 503 ] }, { "type": "text", "bbox": [ 106, 507, 505, 542 ], "lines": [ { "bbox": [ 105, 507, 505, 521 ], "spans": [ { "bbox": [ 105, 507, 455, 521 ], "score": 1.0, "content": "Style-content reconstruction loss. Similar to StarGAN v2, to enforce the generator", "type": "text" }, { "bbox": [ 456, 509, 465, 518 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 465, 507, 505, 521 ], "score": 1.0, "content": "takes the", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 519, 505, 532 ], "spans": [ { "bbox": [ 105, 519, 149, 532 ], "score": 1.0, "content": "style code", "type": "text" }, { "bbox": [ 149, 519, 156, 529 ], "score": 0.64, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 156, 519, 505, 532 ], "score": 1.0, "content": "and the domain specific content ec into consideration during the generation process, we", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 531, 278, 543 ], "spans": [ { "bbox": [ 106, 531, 278, 543 ], "score": 1.0, "content": "employ a style-content reconstruction loss", "type": "text" } ], "index": 35 } ], "index": 34, "bbox_fs": [ 105, 507, 505, 543 ] }, { "type": "interline_equation", "bbox": [ 148, 547, 463, 563 ], "lines": [ { "bbox": [ 148, 547, 463, 563 ], "spans": [ { "bbox": [ 148, 547, 463, 563 ], "score": 0.89, "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { s t y . c o n } } = \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } } \\big [ \\| \\widetilde { \\mathbf { s } } - S _ { E } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) ) \\| _ { 1 } \\big ] + \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } } \\big [ \\| \\widetilde { \\mathbf { c } } - C _ { E } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) ) \\| _ { 1 } \\big ] . } \\end{array}", "type": "interline_equation", "image_path": "ae70d605b176b542cef1b090f291d2c692ecbbaaceb1c0f0b119e790a5f72661.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 148, 547, 463, 563 ], "spans": [], "index": 36 } ] }, { "type": "text", "bbox": [ 106, 568, 505, 603 ], "lines": [ { "bbox": [ 106, 568, 505, 581 ], "spans": [ { "bbox": [ 106, 569, 291, 581 ], "score": 1.0, "content": "This objective urges the style-content encoder", "type": "text" }, { "bbox": [ 291, 569, 300, 579 ], "score": 0.84, "content": "E", "type": "inline_equation" }, { "bbox": [ 300, 569, 342, 581 ], "score": 1.0, "content": "to recover", "type": "text" }, { "bbox": [ 343, 568, 350, 579 ], "score": 0.47, "content": "\\widetilde { \\mathbf { s } }", "type": "inline_equation" }, { "bbox": [ 350, 569, 396, 581 ], "score": 1.0, "content": "and c from", "type": "text" }, { "bbox": [ 396, 568, 437, 581 ], "score": 0.91, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 437, 569, 505, 581 ], "score": 1.0, "content": ". Here, the style-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 579, 506, 593 ], "spans": [ { "bbox": [ 105, 579, 173, 593 ], "score": 1.0, "content": "content encoder", "type": "text" }, { "bbox": [ 173, 580, 182, 590 ], "score": 0.83, "content": "E", "type": "inline_equation" }, { "bbox": [ 183, 579, 506, 593 ], "score": 1.0, "content": "learns a mapping from an image to its style and content domains, which allows", "type": "text" } ], "index": 38 }, { "bbox": [ 107, 590, 426, 603 ], "spans": [ { "bbox": [ 107, 592, 115, 601 ], "score": 0.8, "content": "G", "type": "inline_equation" }, { "bbox": [ 116, 590, 426, 603 ], "score": 1.0, "content": "to synthesize an image with given s and c from reference images at test time.", "type": "text" } ], "index": 39 } ], "index": 38, "bbox_fs": [ 105, 568, 506, 603 ] }, { "type": "text", "bbox": [ 107, 607, 504, 652 ], "lines": [ { "bbox": [ 105, 606, 506, 620 ], "spans": [ { "bbox": [ 105, 606, 427, 620 ], "score": 1.0, "content": "Diversity loss. In order to further boost the diversity of output images from", "type": "text" }, { "bbox": [ 428, 608, 436, 618 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 437, 606, 506, 620 ], "score": 1.0, "content": ", we introduce a", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 619, 505, 631 ], "spans": [ { "bbox": [ 105, 619, 411, 631 ], "score": 1.0, "content": "loss that encourages diversity as follows: for a pair of random latent codes", "type": "text" }, { "bbox": [ 411, 620, 422, 630 ], "score": 0.85, "content": "\\mathbf { z } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 423, 619, 441, 631 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 441, 620, 452, 630 ], "score": 0.87, "content": "\\mathbf { z } _ { 2 }", "type": "inline_equation" }, { "bbox": [ 452, 619, 505, 631 ], "score": 1.0, "content": "we compute", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 630, 504, 642 ], "spans": [ { "bbox": [ 106, 630, 174, 642 ], "score": 0.93, "content": "\\widetilde { \\mathbf { s } } _ { i } , \\widetilde { \\mathbf { c } } _ { i } = M _ { \\widetilde { y } } ( \\mathbf { z } _ { i } )", "type": "inline_equation" }, { "bbox": [ 174, 630, 191, 642 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 191, 630, 234, 642 ], "score": 0.93, "content": "i \\in \\{ 1 , 2 \\}", "type": "inline_equation" }, { "bbox": [ 234, 630, 435, 642 ], "score": 1.0, "content": "and enforce a different outcome of the generator", "type": "text" }, { "bbox": [ 435, 630, 444, 640 ], "score": 0.84, "content": "G", "type": "inline_equation" }, { "bbox": [ 444, 630, 504, 642 ], "score": 1.0, "content": "for differently", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 640, 252, 653 ], "spans": [ { "bbox": [ 105, 640, 252, 653 ], "score": 1.0, "content": "mixed style and content input pairs:", "type": "text" } ], "index": 43 } ], "index": 41.5, "bbox_fs": [ 105, 606, 506, 653 ] }, { "type": "interline_equation", "bbox": [ 182, 656, 428, 706 ], "lines": [ { "bbox": [ 182, 656, 428, 706 ], "spans": [ { "bbox": [ 182, 656, 428, 706 ], "score": 0.93, "content": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { d s } } = \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } } \\left[ \\| G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 1 } , \\widetilde { \\mathbf { c } } _ { 2 } ) - G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 2 } , \\widetilde { \\mathbf { c } } _ { 1 } ) \\| _ { 1 } \\right] } \\\\ & { \\quad \\quad + \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } } \\left[ \\| G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 1 } , \\widetilde { \\mathbf { c } } _ { 1 } ) - G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { 2 } , \\widetilde { \\mathbf { c } } _ { 2 } ) \\| _ { 1 } \\right] } \\\\ & { \\quad \\quad + \\sum _ { m , n , o } \\left[ \\mathbb { E } _ { \\mathbf { x } , \\widetilde { y } , \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } } \\left[ \\| G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { m } , \\widetilde { \\mathbf { c } } _ { n } ) - G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } _ { o } , \\widetilde { \\mathbf { c } } _ { o } ) \\| _ { 1 } \\right] \\right] , } \\end{array}", "type": "interline_equation", "image_path": "af2defdd3af95b358356d3f45b7f10b9e2ba3e2f9851f7366b219e39a10c234c.jpg" } ] } ], "index": 45, "virtual_lines": [ { "bbox": [ 182, 656, 428, 672.6666666666666 ], "spans": [], "index": 44 }, { "bbox": [ 182, 672.6666666666666, 428, 689.3333333333333 ], "spans": [], "index": 45 }, { "bbox": [ 182, 689.3333333333333, 428, 705.9999999999999 ], "spans": [], "index": 46 } ] }, { "type": "text", "bbox": [ 108, 709, 503, 732 ], "lines": [ { "bbox": [ 106, 709, 505, 722 ], "spans": [ { "bbox": [ 106, 709, 134, 722 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 709, 228, 722 ], "score": 0.92, "content": "m , n \\in \\{ 1 , 2 | m \\neq n \\}", "type": "inline_equation" }, { "bbox": [ 229, 709, 249, 722 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 249, 710, 294, 722 ], "score": 0.92, "content": "o \\in \\{ 1 , 2 \\}", "type": "inline_equation" }, { "bbox": [ 295, 709, 444, 722 ], "score": 1.0, "content": ". Driven by this term, the generator", "type": "text" }, { "bbox": [ 445, 710, 454, 720 ], "score": 0.82, "content": "G", "type": "inline_equation" }, { "bbox": [ 454, 709, 505, 722 ], "score": 1.0, "content": "is forced to", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 719, 505, 734 ], "spans": [ { "bbox": [ 106, 719, 505, 734 ], "score": 1.0, "content": "discover meaningful style features and contents that eventually lead to diversity in generated images.", "type": "text" } ], "index": 48 } ], "index": 47.5, "bbox_fs": [ 106, 709, 505, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 504, 105 ], "lines": [ { "bbox": [ 105, 81, 506, 96 ], "spans": [ { "bbox": [ 105, 81, 219, 96 ], "score": 1.0, "content": "We ignore the denominator", "type": "text" }, { "bbox": [ 220, 82, 266, 95 ], "score": 0.93, "content": "{ \\left\\| { \\bf z } _ { 1 } - { \\bf z } _ { 2 } \\right\\| } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 267, 81, 506, 96 ], "score": 1.0, "content": "of the original diversity loss (Mao et al., 2019a) for stable", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 216, 105 ], "spans": [ { "bbox": [ 105, 93, 216, 105 ], "score": 1.0, "content": "training as in StarGAN v2.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 105, 110, 503, 133 ], "lines": [ { "bbox": [ 106, 110, 504, 122 ], "spans": [ { "bbox": [ 106, 111, 366, 122 ], "score": 1.0, "content": "Cycle consistency loss. To ensure that the generated image", "type": "text" }, { "bbox": [ 367, 110, 407, 122 ], "score": 0.93, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 408, 111, 504, 122 ], "score": 1.0, "content": "preserves the domain-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 121, 497, 134 ], "spans": [ { "bbox": [ 105, 121, 258, 134 ], "score": 1.0, "content": "invariant properties of its input image", "type": "text" }, { "bbox": [ 258, 124, 266, 131 ], "score": 0.58, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 266, 121, 497, 134 ], "score": 1.0, "content": ", we impose the cycle consistency loss (Zhu et al., 2017a)", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "interline_equation", "bbox": [ 212, 137, 397, 152 ], "lines": [ { "bbox": [ 212, 137, 397, 152 ], "spans": [ { "bbox": [ 212, 137, 397, 152 ], "score": 0.93, "content": "\\mathcal { L } _ { \\mathrm { c y c } } = \\mathbb { E } _ { { \\mathbf { x } } , y , \\widetilde { y } , { \\mathbf { z } } } \\big [ | | { \\mathbf { x } } - G ( G ( { \\mathbf { x } } , \\widetilde { { \\mathbf { s } } } , \\widetilde { { \\mathbf { c } } } ) , \\widehat { { \\mathbf { s } } } , \\widehat { { \\mathbf { c } } } ) | | _ { 1 } \\big ] ,", "type": "interline_equation", "image_path": "8a28d37d9ae5394b57315d9c8d4765defe93168ce7c89ec9750064fee390377d.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 212, 137, 397, 152 ], "spans": [], "index": 4 } ] }, { "type": "text", "bbox": [ 106, 156, 505, 201 ], "lines": [ { "bbox": [ 105, 155, 506, 170 ], "spans": [ { "bbox": [ 105, 155, 133, 170 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 156, 190, 169 ], "score": 0.93, "content": "\\hat { \\bf s } , \\hat { \\bf c } = E _ { y } ( { \\bf x } )", "type": "inline_equation" }, { "bbox": [ 191, 155, 493, 170 ], "score": 1.0, "content": "is the extracted style code and domain specific content of the input image", "type": "text" }, { "bbox": [ 494, 159, 501, 167 ], "score": 0.37, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 502, 155, 506, 170 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 168, 505, 180 ], "spans": [ { "bbox": [ 106, 168, 124, 180 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 124, 169, 131, 179 ], "score": 0.81, "content": "y", "type": "inline_equation" }, { "bbox": [ 131, 168, 237, 180 ], "score": 1.0, "content": "is the original domain of", "type": "text" }, { "bbox": [ 237, 169, 245, 177 ], "score": 0.6, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 245, 168, 428, 180 ], "score": 1.0, "content": ". By learning to reconstruct the input image", "type": "text" }, { "bbox": [ 428, 169, 436, 178 ], "score": 0.56, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 436, 168, 505, 180 ], "score": 1.0, "content": "with given style", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 179, 505, 191 ], "spans": [ { "bbox": [ 106, 179, 248, 191 ], "score": 1.0, "content": "code ˆs and content cˆ, the generator", "type": "text" }, { "bbox": [ 249, 179, 258, 188 ], "score": 0.84, "content": "G", "type": "inline_equation" }, { "bbox": [ 258, 179, 505, 191 ], "score": 1.0, "content": "is then further encouraged to disentangle the background, the", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 190, 280, 201 ], "spans": [ { "bbox": [ 106, 190, 280, 201 ], "score": 1.0, "content": "domain specific content and the style code.", "type": "text" } ], "index": 8 } ], "index": 6.5 }, { "type": "text", "bbox": [ 106, 206, 505, 240 ], "lines": [ { "bbox": [ 106, 206, 505, 218 ], "spans": [ { "bbox": [ 106, 206, 505, 218 ], "score": 1.0, "content": "Content consistency loss. Besides the cycle consistency loss, we apply another constraint to en-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 217, 504, 229 ], "spans": [ { "bbox": [ 106, 217, 320, 229 ], "score": 1.0, "content": "force that the detached domain specific content from", "type": "text" }, { "bbox": [ 320, 218, 329, 227 ], "score": 0.83, "content": "G", "type": "inline_equation" }, { "bbox": [ 330, 217, 495, 229 ], "score": 1.0, "content": "is consistent with the one retrieved from", "type": "text" }, { "bbox": [ 495, 218, 504, 227 ], "score": 0.81, "content": "E", "type": "inline_equation" } ], "index": 10 }, { "bbox": [ 105, 228, 160, 242 ], "spans": [ { "bbox": [ 105, 228, 160, 242 ], "score": 1.0, "content": "according to", "type": "text" } ], "index": 11 } ], "index": 10 }, { "type": "interline_equation", "bbox": [ 149, 244, 461, 259 ], "lines": [ { "bbox": [ 149, 244, 461, 259 ], "spans": [ { "bbox": [ 149, 244, 461, 259 ], "score": 0.89, "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { c o n . c y c } } = \\mathbb { E } _ { \\mathbf { x } , y , \\widetilde { y } , \\mathbf { z } } \\left[ \\left\\| F G _ { G } ( \\mathbf { x } ) - \\widehat { \\mathbf { c } } \\right\\| _ { 1 } \\right] + \\mathbb { E } _ { \\mathbf { x } , y , \\widetilde { y } , \\mathbf { z } } \\left[ \\left\\| F G _ { G } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) ) - \\widetilde { \\mathbf { c } } \\right\\| _ { 1 } \\right] , } \\end{array}", "type": "interline_equation", "image_path": "16799b09efe3daa473a44c7377c649313946b26779783a12f869bc3fcbfa421e.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 149, 244, 461, 259 ], "spans": [], "index": 12 } ] }, { "type": "text", "bbox": [ 108, 262, 504, 286 ], "lines": [ { "bbox": [ 105, 262, 505, 276 ], "spans": [ { "bbox": [ 105, 262, 134, 276 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 262, 280, 276 ], "score": 0.31, "content": "\\hat { \\mathbf { c } } = E _ { y } ( \\mathbf { x } ) , \\widetilde { \\mathbf { c } } = E _ { \\widetilde { y } } ( \\mathbf { x } ) , F G _ { G } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 280, 262, 299, 276 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 300, 263, 370, 275 ], "score": 0.92, "content": "F G _ { G } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) )", "type": "inline_equation" }, { "bbox": [ 370, 262, 505, 276 ], "score": 1.0, "content": "are the pop-out domain specific", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 275, 401, 288 ], "spans": [ { "bbox": [ 106, 275, 210, 288 ], "score": 1.0, "content": "content from input image", "type": "text" }, { "bbox": [ 210, 277, 218, 284 ], "score": 0.48, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 218, 275, 303, 288 ], "score": 1.0, "content": "and generated image", "type": "text" }, { "bbox": [ 304, 275, 344, 286 ], "score": 0.92, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 344, 275, 401, 288 ], "score": 1.0, "content": ", respectively.", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 105, 291, 504, 313 ], "lines": [ { "bbox": [ 105, 290, 505, 304 ], "spans": [ { "bbox": [ 105, 290, 505, 304 ], "score": 1.0, "content": "Classification losses. We employ two classification losses: the first one is the foreground content", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 301, 180, 314 ], "spans": [ { "bbox": [ 106, 301, 180, 314 ], "score": 1.0, "content": "classification loss", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "interline_equation", "bbox": [ 152, 318, 457, 334 ], "lines": [ { "bbox": [ 152, 318, 457, 334 ], "spans": [ { "bbox": [ 152, 318, 457, 334 ], "score": 0.88, "content": "\\mathcal { L } _ { \\mathrm { F G . c l s } } = \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { r e a l } } , y } \\Big [ - \\log D _ { \\mathrm { F G . c l s } } \\big ( y | \\mathbf { x } _ { \\mathrm { r e a l } } \\big ) \\Big ] + \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { f a k e } } , \\widetilde { y } } \\Big [ - \\log D _ { \\mathrm { F G . c l s } } \\big ( \\widetilde { y } | \\mathbf { x } _ { \\mathrm { f a k e } } \\big ) \\Big ] \\ ,", "type": "interline_equation", "image_path": "6712ae0ca7d1db4598667f0ee1d11db11e23257990d2e793428b4f8a10453059.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 152, 318, 457, 334 ], "spans": [], "index": 17 } ] }, { "type": "text", "bbox": [ 107, 337, 504, 360 ], "lines": [ { "bbox": [ 106, 337, 504, 350 ], "spans": [ { "bbox": [ 106, 337, 504, 350 ], "score": 1.0, "content": "which aims to ensure that the domain specific content is properly encoded and carries enough infor-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 349, 444, 360 ], "spans": [ { "bbox": [ 106, 349, 444, 360 ], "score": 1.0, "content": "mation from the target domain. The second one is the background classification loss", "type": "text" } ], "index": 19 } ], "index": 18.5 }, { "type": "interline_equation", "bbox": [ 151, 363, 458, 379 ], "lines": [ { "bbox": [ 151, 363, 458, 379 ], "spans": [ { "bbox": [ 151, 363, 458, 379 ], "score": 0.89, "content": "\\mathcal { L } _ { \\mathrm { B G } . \\mathrm { c l s } } = \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { r e a l } } , p } \\big [ - \\log D _ { \\mathrm { B G } . \\mathrm { c l s } } ( p | \\mathbf { x } _ { \\mathrm { r e a l } } ) \\big ] + \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { f a k e } } , p } \\big [ - \\log D _ { \\mathrm { B G } . \\mathrm { c l s } } ( p | \\mathbf { x } _ { \\mathrm { f a k e } } ) \\big ] \\ ,", "type": "interline_equation", "image_path": "115c772df39b6fd258d86dbe2a48df8fb2caeffe1cc1c194cbf32120e55fcba7.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 151, 363, 458, 379 ], "spans": [], "index": 20 } ] }, { "type": "text", "bbox": [ 107, 383, 505, 417 ], "lines": [ { "bbox": [ 105, 382, 505, 397 ], "spans": [ { "bbox": [ 105, 382, 133, 397 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 385, 141, 395 ], "score": 0.81, "content": "p", "type": "inline_equation" }, { "bbox": [ 141, 382, 311, 397 ], "score": 1.0, "content": "is the corresponding background type of", "type": "text" }, { "bbox": [ 311, 385, 329, 394 ], "score": 0.88, "content": "\\mathbf { x } _ { \\mathrm { r e a l } }", "type": "inline_equation" }, { "bbox": [ 330, 382, 349, 397 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 349, 385, 369, 395 ], "score": 0.89, "content": "\\mathbf { x } _ { \\mathrm { f a k e } }", "type": "inline_equation" }, { "bbox": [ 369, 382, 505, 397 ], "score": 1.0, "content": ". With the help of this objective,", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 395, 504, 407 ], "spans": [ { "bbox": [ 106, 395, 163, 407 ], "score": 1.0, "content": "the generator", "type": "text" }, { "bbox": [ 163, 395, 172, 405 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 172, 395, 470, 407 ], "score": 1.0, "content": "learns to preserve the domain-invariant characteristics of its input image", "type": "text" }, { "bbox": [ 471, 396, 479, 405 ], "score": 0.67, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 479, 395, 504, 407 ], "score": 1.0, "content": "while", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 405, 303, 418 ], "spans": [ { "bbox": [ 106, 405, 303, 418 ], "score": 1.0, "content": "dissociating the foreground domain specific part.", "type": "text" } ], "index": 23 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 422, 372, 434 ], "lines": [ { "bbox": [ 105, 421, 373, 435 ], "spans": [ { "bbox": [ 105, 421, 373, 435 ], "score": 1.0, "content": "Full objective. Our full objective functions can be summarized as", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "interline_equation", "bbox": [ 178, 437, 433, 472 ], "lines": [ { "bbox": [ 178, 437, 433, 472 ], "spans": [ { "bbox": [ 178, 437, 433, 472 ], "score": 0.87, "content": "\\begin{array} { r l } { \\underset { G , F , E } { \\operatorname* { m i n } } \\underset { D } { \\operatorname* { m a x } } } & { \\mathcal { L } _ { \\mathrm { a d v } } + \\lambda _ { \\mathrm { s t y } . \\mathrm { c o n } } \\mathcal { L } _ { \\mathrm { s t y } . \\mathrm { c o n } } - \\lambda _ { \\mathrm { d s } } \\mathcal { L } _ { \\mathrm { d s } } + \\lambda _ { \\mathrm { c y c } } \\mathcal { L } _ { \\mathrm { c y c } } + } \\\\ & { \\lambda _ { \\mathrm { c o n . c y c } } \\mathcal { L } _ { \\mathrm { c o n . c y c } } + \\lambda _ { \\mathrm { F G . c l s } } \\mathcal { L } _ { \\mathrm { F G . c l s } } + \\lambda _ { \\mathrm { B G . c l s } } \\mathcal { L } _ { \\mathrm { B G . c l s } } \\ , } \\end{array}", "type": "interline_equation", "image_path": "05fe2b4c8716195856bd4bb70ae9fafb48b7dd38578c2cce55917eb22af3c280.jpg" } ] } ], "index": 26, "virtual_lines": [ { "bbox": [ 178, 437, 433, 448.6666666666667 ], "spans": [], "index": 25 }, { "bbox": [ 178, 448.6666666666667, 433, 460.33333333333337 ], "spans": [], "index": 26 }, { "bbox": [ 178, 460.33333333333337, 433, 472.00000000000006 ], "spans": [], "index": 27 } ] }, { "type": "text", "bbox": [ 106, 474, 455, 488 ], "lines": [ { "bbox": [ 105, 473, 458, 489 ], "spans": [ { "bbox": [ 105, 473, 133, 489 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 475, 149, 487 ], "score": 0.32, "content": "\\lambda _ { \\mathrm { s t y } }", "type": "inline_equation" }, { "bbox": [ 149, 473, 153, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 153, 475, 166, 487 ], "score": 0.35, "content": "\\lambda _ { \\mathrm { d s } }", "type": "inline_equation" }, { "bbox": [ 167, 473, 170, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 170, 475, 187, 488 ], "score": 0.49, "content": "\\lambda _ { \\mathrm { c y c } }", "type": "inline_equation" }, { "bbox": [ 188, 473, 193, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 194, 475, 223, 487 ], "score": 0.73, "content": "\\lambda _ { \\mathrm { c o n \\mathrm { { - } c y c } } }", "type": "inline_equation" }, { "bbox": [ 223, 473, 227, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 227, 475, 254, 487 ], "score": 0.85, "content": "\\lambda _ { \\mathrm { F G \\mathrm { - } c l s } }", "type": "inline_equation" }, { "bbox": [ 254, 473, 272, 489 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 272, 475, 299, 487 ], "score": 0.91, "content": "\\lambda _ { \\mathrm { B G \\mathrm { { - } c l s } } }", "type": "inline_equation" }, { "bbox": [ 300, 473, 458, 489 ], "score": 1.0, "content": "are the hyperparameters for each term.", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "title", "bbox": [ 107, 502, 200, 515 ], "lines": [ { "bbox": [ 105, 502, 201, 516 ], "spans": [ { "bbox": [ 105, 502, 201, 516 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 108, 527, 504, 560 ], "lines": [ { "bbox": [ 105, 525, 505, 540 ], "spans": [ { "bbox": [ 105, 525, 505, 540 ], "score": 1.0, "content": "We evaluated the images generated by DT-GAN through a series of experiments both quantitatively", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 538, 505, 550 ], "spans": [ { "bbox": [ 106, 538, 505, 550 ], "score": 1.0, "content": "and qualitatively. Finally, we demonstrate the benefits of our generated images when being used as", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 549, 368, 561 ], "spans": [ { "bbox": [ 106, 549, 368, 561 ], "score": 1.0, "content": "data augmentation for a defect classification task on limited data.", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 106, 565, 505, 665 ], "lines": [ { "bbox": [ 106, 566, 505, 577 ], "spans": [ { "bbox": [ 106, 566, 505, 577 ], "score": 1.0, "content": "Dataset. All experiments were performed on a real industrial dataset: a Surface Defect Inspection", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 577, 504, 589 ], "spans": [ { "bbox": [ 106, 577, 504, 589 ], "score": 1.0, "content": "(SDI) dataset that contains three different kinds of products from production lines and samples from", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 588, 505, 600 ], "spans": [ { "bbox": [ 106, 588, 505, 600 ], "score": 1.0, "content": "each product are classified into three mutually exclusive classes: Normal, Scratch and Spot.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 598, 505, 612 ], "spans": [ { "bbox": [ 105, 598, 505, 612 ], "score": 1.0, "content": "All of the images are grayscale. Detailed statistics of the dataset are summarized in Appendix", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 609, 506, 622 ], "spans": [ { "bbox": [ 105, 609, 506, 622 ], "score": 1.0, "content": "A. Note that only the training set was used in GAN training, the test set was left untouched for", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 621, 504, 634 ], "spans": [ { "bbox": [ 106, 621, 459, 634 ], "score": 1.0, "content": "final evaluation in classifier training. For a fair comparison, all images were resized to", "type": "text" }, { "bbox": [ 460, 621, 504, 631 ], "score": 0.89, "content": "1 2 8 \\times 1 2 8", "type": "inline_equation" } ], "index": 38 }, { "bbox": [ 106, 632, 505, 644 ], "spans": [ { "bbox": [ 106, 632, 505, 644 ], "score": 1.0, "content": "resolution for both GAN training and classifier training, which was also the highest resolution used", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 642, 505, 656 ], "spans": [ { "bbox": [ 105, 642, 505, 656 ], "score": 1.0, "content": "in the baselines for image generation. For comparison, we also conducted experiments on the widely", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 653, 439, 666 ], "spans": [ { "bbox": [ 105, 653, 439, 666 ], "score": 1.0, "content": "used MVTec Anomaly Detection dataset (Bergmann et al., 2019) in Appendix E.4.", "type": "text" } ], "index": 41 } ], "index": 37 }, { "type": "title", "bbox": [ 108, 678, 225, 689 ], "lines": [ { "bbox": [ 105, 677, 227, 691 ], "spans": [ { "bbox": [ 105, 677, 227, 691 ], "score": 1.0, "content": "4.1 DEFECT GENERATION", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Baselines. As discussed in Section 3, DT-GAN can either use the mapping network to randomly", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "generate styles and defects, or it can use the style-content encoder to extract both from reference", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 720, 454, 734 ], "spans": [ { "bbox": [ 105, 720, 454, 734 ], "score": 1.0, "content": "images. We refer to these cases as ‘latent-guided’ and ‘reference-guided’, respectively.", "type": "text" } ], "index": 45 } ], "index": 44 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 26, 308, 38 ], "lines": [ { "bbox": [ 107, 25, 308, 39 ], "spans": [ { "bbox": [ 107, 25, 308, 39 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 763 ], "spans": [ { "bbox": [ 302, 750, 309, 763 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 504, 105 ], "lines": [ { "bbox": [ 105, 81, 506, 96 ], "spans": [ { "bbox": [ 105, 81, 219, 96 ], "score": 1.0, "content": "We ignore the denominator", "type": "text" }, { "bbox": [ 220, 82, 266, 95 ], "score": 0.93, "content": "{ \\left\\| { \\bf z } _ { 1 } - { \\bf z } _ { 2 } \\right\\| } _ { 1 }", "type": "inline_equation" }, { "bbox": [ 267, 81, 506, 96 ], "score": 1.0, "content": "of the original diversity loss (Mao et al., 2019a) for stable", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 216, 105 ], "spans": [ { "bbox": [ 105, 93, 216, 105 ], "score": 1.0, "content": "training as in StarGAN v2.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 81, 506, 105 ] }, { "type": "text", "bbox": [ 105, 110, 503, 133 ], "lines": [ { "bbox": [ 106, 110, 504, 122 ], "spans": [ { "bbox": [ 106, 111, 366, 122 ], "score": 1.0, "content": "Cycle consistency loss. To ensure that the generated image", "type": "text" }, { "bbox": [ 367, 110, 407, 122 ], "score": 0.93, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 408, 111, 504, 122 ], "score": 1.0, "content": "preserves the domain-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 121, 497, 134 ], "spans": [ { "bbox": [ 105, 121, 258, 134 ], "score": 1.0, "content": "invariant properties of its input image", "type": "text" }, { "bbox": [ 258, 124, 266, 131 ], "score": 0.58, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 266, 121, 497, 134 ], "score": 1.0, "content": ", we impose the cycle consistency loss (Zhu et al., 2017a)", "type": "text" } ], "index": 3 } ], "index": 2.5, "bbox_fs": [ 105, 110, 504, 134 ] }, { "type": "interline_equation", "bbox": [ 212, 137, 397, 152 ], "lines": [ { "bbox": [ 212, 137, 397, 152 ], "spans": [ { "bbox": [ 212, 137, 397, 152 ], "score": 0.93, "content": "\\mathcal { L } _ { \\mathrm { c y c } } = \\mathbb { E } _ { { \\mathbf { x } } , y , \\widetilde { y } , { \\mathbf { z } } } \\big [ | | { \\mathbf { x } } - G ( G ( { \\mathbf { x } } , \\widetilde { { \\mathbf { s } } } , \\widetilde { { \\mathbf { c } } } ) , \\widehat { { \\mathbf { s } } } , \\widehat { { \\mathbf { c } } } ) | | _ { 1 } \\big ] ,", "type": "interline_equation", "image_path": "8a28d37d9ae5394b57315d9c8d4765defe93168ce7c89ec9750064fee390377d.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 212, 137, 397, 152 ], "spans": [], "index": 4 } ] }, { "type": "text", "bbox": [ 106, 156, 505, 201 ], "lines": [ { "bbox": [ 105, 155, 506, 170 ], "spans": [ { "bbox": [ 105, 155, 133, 170 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 156, 190, 169 ], "score": 0.93, "content": "\\hat { \\bf s } , \\hat { \\bf c } = E _ { y } ( { \\bf x } )", "type": "inline_equation" }, { "bbox": [ 191, 155, 493, 170 ], "score": 1.0, "content": "is the extracted style code and domain specific content of the input image", "type": "text" }, { "bbox": [ 494, 159, 501, 167 ], "score": 0.37, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 502, 155, 506, 170 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 168, 505, 180 ], "spans": [ { "bbox": [ 106, 168, 124, 180 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 124, 169, 131, 179 ], "score": 0.81, "content": "y", "type": "inline_equation" }, { "bbox": [ 131, 168, 237, 180 ], "score": 1.0, "content": "is the original domain of", "type": "text" }, { "bbox": [ 237, 169, 245, 177 ], "score": 0.6, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 245, 168, 428, 180 ], "score": 1.0, "content": ". By learning to reconstruct the input image", "type": "text" }, { "bbox": [ 428, 169, 436, 178 ], "score": 0.56, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 436, 168, 505, 180 ], "score": 1.0, "content": "with given style", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 179, 505, 191 ], "spans": [ { "bbox": [ 106, 179, 248, 191 ], "score": 1.0, "content": "code ˆs and content cˆ, the generator", "type": "text" }, { "bbox": [ 249, 179, 258, 188 ], "score": 0.84, "content": "G", "type": "inline_equation" }, { "bbox": [ 258, 179, 505, 191 ], "score": 1.0, "content": "is then further encouraged to disentangle the background, the", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 190, 280, 201 ], "spans": [ { "bbox": [ 106, 190, 280, 201 ], "score": 1.0, "content": "domain specific content and the style code.", "type": "text" } ], "index": 8 } ], "index": 6.5, "bbox_fs": [ 105, 155, 506, 201 ] }, { "type": "text", "bbox": [ 106, 206, 505, 240 ], "lines": [ { "bbox": [ 106, 206, 505, 218 ], "spans": [ { "bbox": [ 106, 206, 505, 218 ], "score": 1.0, "content": "Content consistency loss. Besides the cycle consistency loss, we apply another constraint to en-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 217, 504, 229 ], "spans": [ { "bbox": [ 106, 217, 320, 229 ], "score": 1.0, "content": "force that the detached domain specific content from", "type": "text" }, { "bbox": [ 320, 218, 329, 227 ], "score": 0.83, "content": "G", "type": "inline_equation" }, { "bbox": [ 330, 217, 495, 229 ], "score": 1.0, "content": "is consistent with the one retrieved from", "type": "text" }, { "bbox": [ 495, 218, 504, 227 ], "score": 0.81, "content": "E", "type": "inline_equation" } ], "index": 10 }, { "bbox": [ 105, 228, 160, 242 ], "spans": [ { "bbox": [ 105, 228, 160, 242 ], "score": 1.0, "content": "according to", "type": "text" } ], "index": 11 } ], "index": 10, "bbox_fs": [ 105, 206, 505, 242 ] }, { "type": "interline_equation", "bbox": [ 149, 244, 461, 259 ], "lines": [ { "bbox": [ 149, 244, 461, 259 ], "spans": [ { "bbox": [ 149, 244, 461, 259 ], "score": 0.89, "content": "\\begin{array} { r } { \\mathcal { L } _ { \\mathrm { c o n . c y c } } = \\mathbb { E } _ { \\mathbf { x } , y , \\widetilde { y } , \\mathbf { z } } \\left[ \\left\\| F G _ { G } ( \\mathbf { x } ) - \\widehat { \\mathbf { c } } \\right\\| _ { 1 } \\right] + \\mathbb { E } _ { \\mathbf { x } , y , \\widetilde { y } , \\mathbf { z } } \\left[ \\left\\| F G _ { G } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) ) - \\widetilde { \\mathbf { c } } \\right\\| _ { 1 } \\right] , } \\end{array}", "type": "interline_equation", "image_path": "16799b09efe3daa473a44c7377c649313946b26779783a12f869bc3fcbfa421e.jpg" } ] } ], "index": 12, "virtual_lines": [ { "bbox": [ 149, 244, 461, 259 ], "spans": [], "index": 12 } ] }, { "type": "text", "bbox": [ 108, 262, 504, 286 ], "lines": [ { "bbox": [ 105, 262, 505, 276 ], "spans": [ { "bbox": [ 105, 262, 134, 276 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 262, 280, 276 ], "score": 0.31, "content": "\\hat { \\mathbf { c } } = E _ { y } ( \\mathbf { x } ) , \\widetilde { \\mathbf { c } } = E _ { \\widetilde { y } } ( \\mathbf { x } ) , F G _ { G } ( \\mathbf { x } )", "type": "inline_equation" }, { "bbox": [ 280, 262, 299, 276 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 300, 263, 370, 275 ], "score": 0.92, "content": "F G _ { G } ( G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } ) )", "type": "inline_equation" }, { "bbox": [ 370, 262, 505, 276 ], "score": 1.0, "content": "are the pop-out domain specific", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 275, 401, 288 ], "spans": [ { "bbox": [ 106, 275, 210, 288 ], "score": 1.0, "content": "content from input image", "type": "text" }, { "bbox": [ 210, 277, 218, 284 ], "score": 0.48, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 218, 275, 303, 288 ], "score": 1.0, "content": "and generated image", "type": "text" }, { "bbox": [ 304, 275, 344, 286 ], "score": 0.92, "content": "G ( \\mathbf { x } , \\widetilde { \\mathbf { s } } , \\widetilde { \\mathbf { c } } )", "type": "inline_equation" }, { "bbox": [ 344, 275, 401, 288 ], "score": 1.0, "content": ", respectively.", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 105, 262, 505, 288 ] }, { "type": "text", "bbox": [ 105, 291, 504, 313 ], "lines": [ { "bbox": [ 105, 290, 505, 304 ], "spans": [ { "bbox": [ 105, 290, 505, 304 ], "score": 1.0, "content": "Classification losses. We employ two classification losses: the first one is the foreground content", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 301, 180, 314 ], "spans": [ { "bbox": [ 106, 301, 180, 314 ], "score": 1.0, "content": "classification loss", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 105, 290, 505, 314 ] }, { "type": "interline_equation", "bbox": [ 152, 318, 457, 334 ], "lines": [ { "bbox": [ 152, 318, 457, 334 ], "spans": [ { "bbox": [ 152, 318, 457, 334 ], "score": 0.88, "content": "\\mathcal { L } _ { \\mathrm { F G . c l s } } = \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { r e a l } } , y } \\Big [ - \\log D _ { \\mathrm { F G . c l s } } \\big ( y | \\mathbf { x } _ { \\mathrm { r e a l } } \\big ) \\Big ] + \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { f a k e } } , \\widetilde { y } } \\Big [ - \\log D _ { \\mathrm { F G . c l s } } \\big ( \\widetilde { y } | \\mathbf { x } _ { \\mathrm { f a k e } } \\big ) \\Big ] \\ ,", "type": "interline_equation", "image_path": "6712ae0ca7d1db4598667f0ee1d11db11e23257990d2e793428b4f8a10453059.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 152, 318, 457, 334 ], "spans": [], "index": 17 } ] }, { "type": "text", "bbox": [ 107, 337, 504, 360 ], "lines": [ { "bbox": [ 106, 337, 504, 350 ], "spans": [ { "bbox": [ 106, 337, 504, 350 ], "score": 1.0, "content": "which aims to ensure that the domain specific content is properly encoded and carries enough infor-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 349, 444, 360 ], "spans": [ { "bbox": [ 106, 349, 444, 360 ], "score": 1.0, "content": "mation from the target domain. The second one is the background classification loss", "type": "text" } ], "index": 19 } ], "index": 18.5, "bbox_fs": [ 106, 337, 504, 360 ] }, { "type": "interline_equation", "bbox": [ 151, 363, 458, 379 ], "lines": [ { "bbox": [ 151, 363, 458, 379 ], "spans": [ { "bbox": [ 151, 363, 458, 379 ], "score": 0.89, "content": "\\mathcal { L } _ { \\mathrm { B G } . \\mathrm { c l s } } = \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { r e a l } } , p } \\big [ - \\log D _ { \\mathrm { B G } . \\mathrm { c l s } } ( p | \\mathbf { x } _ { \\mathrm { r e a l } } ) \\big ] + \\mathbb { E } _ { \\mathbf { x } _ { \\mathrm { f a k e } } , p } \\big [ - \\log D _ { \\mathrm { B G } . \\mathrm { c l s } } ( p | \\mathbf { x } _ { \\mathrm { f a k e } } ) \\big ] \\ ,", "type": "interline_equation", "image_path": "115c772df39b6fd258d86dbe2a48df8fb2caeffe1cc1c194cbf32120e55fcba7.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 151, 363, 458, 379 ], "spans": [], "index": 20 } ] }, { "type": "text", "bbox": [ 107, 383, 505, 417 ], "lines": [ { "bbox": [ 105, 382, 505, 397 ], "spans": [ { "bbox": [ 105, 382, 133, 397 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 385, 141, 395 ], "score": 0.81, "content": "p", "type": "inline_equation" }, { "bbox": [ 141, 382, 311, 397 ], "score": 1.0, "content": "is the corresponding background type of", "type": "text" }, { "bbox": [ 311, 385, 329, 394 ], "score": 0.88, "content": "\\mathbf { x } _ { \\mathrm { r e a l } }", "type": "inline_equation" }, { "bbox": [ 330, 382, 349, 397 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 349, 385, 369, 395 ], "score": 0.89, "content": "\\mathbf { x } _ { \\mathrm { f a k e } }", "type": "inline_equation" }, { "bbox": [ 369, 382, 505, 397 ], "score": 1.0, "content": ". With the help of this objective,", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 395, 504, 407 ], "spans": [ { "bbox": [ 106, 395, 163, 407 ], "score": 1.0, "content": "the generator", "type": "text" }, { "bbox": [ 163, 395, 172, 405 ], "score": 0.81, "content": "G", "type": "inline_equation" }, { "bbox": [ 172, 395, 470, 407 ], "score": 1.0, "content": "learns to preserve the domain-invariant characteristics of its input image", "type": "text" }, { "bbox": [ 471, 396, 479, 405 ], "score": 0.67, "content": "\\mathbf { x }", "type": "inline_equation" }, { "bbox": [ 479, 395, 504, 407 ], "score": 1.0, "content": "while", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 405, 303, 418 ], "spans": [ { "bbox": [ 106, 405, 303, 418 ], "score": 1.0, "content": "dissociating the foreground domain specific part.", "type": "text" } ], "index": 23 } ], "index": 22, "bbox_fs": [ 105, 382, 505, 418 ] }, { "type": "text", "bbox": [ 107, 422, 372, 434 ], "lines": [ { "bbox": [ 105, 421, 373, 435 ], "spans": [ { "bbox": [ 105, 421, 373, 435 ], "score": 1.0, "content": "Full objective. Our full objective functions can be summarized as", "type": "text" } ], "index": 24 } ], "index": 24, "bbox_fs": [ 105, 421, 373, 435 ] }, { "type": "interline_equation", "bbox": [ 178, 437, 433, 472 ], "lines": [ { "bbox": [ 178, 437, 433, 472 ], "spans": [ { "bbox": [ 178, 437, 433, 472 ], "score": 0.87, "content": "\\begin{array} { r l } { \\underset { G , F , E } { \\operatorname* { m i n } } \\underset { D } { \\operatorname* { m a x } } } & { \\mathcal { L } _ { \\mathrm { a d v } } + \\lambda _ { \\mathrm { s t y } . \\mathrm { c o n } } \\mathcal { L } _ { \\mathrm { s t y } . \\mathrm { c o n } } - \\lambda _ { \\mathrm { d s } } \\mathcal { L } _ { \\mathrm { d s } } + \\lambda _ { \\mathrm { c y c } } \\mathcal { L } _ { \\mathrm { c y c } } + } \\\\ & { \\lambda _ { \\mathrm { c o n . c y c } } \\mathcal { L } _ { \\mathrm { c o n . c y c } } + \\lambda _ { \\mathrm { F G . c l s } } \\mathcal { L } _ { \\mathrm { F G . c l s } } + \\lambda _ { \\mathrm { B G . c l s } } \\mathcal { L } _ { \\mathrm { B G . c l s } } \\ , } \\end{array}", "type": "interline_equation", "image_path": "05fe2b4c8716195856bd4bb70ae9fafb48b7dd38578c2cce55917eb22af3c280.jpg" } ] } ], "index": 26, "virtual_lines": [ { "bbox": [ 178, 437, 433, 448.6666666666667 ], "spans": [], "index": 25 }, { "bbox": [ 178, 448.6666666666667, 433, 460.33333333333337 ], "spans": [], "index": 26 }, { "bbox": [ 178, 460.33333333333337, 433, 472.00000000000006 ], "spans": [], "index": 27 } ] }, { "type": "text", "bbox": [ 106, 474, 455, 488 ], "lines": [ { "bbox": [ 105, 473, 458, 489 ], "spans": [ { "bbox": [ 105, 473, 133, 489 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 133, 475, 149, 487 ], "score": 0.32, "content": "\\lambda _ { \\mathrm { s t y } }", "type": "inline_equation" }, { "bbox": [ 149, 473, 153, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 153, 475, 166, 487 ], "score": 0.35, "content": "\\lambda _ { \\mathrm { d s } }", "type": "inline_equation" }, { "bbox": [ 167, 473, 170, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 170, 475, 187, 488 ], "score": 0.49, "content": "\\lambda _ { \\mathrm { c y c } }", "type": "inline_equation" }, { "bbox": [ 188, 473, 193, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 194, 475, 223, 487 ], "score": 0.73, "content": "\\lambda _ { \\mathrm { c o n \\mathrm { { - } c y c } } }", "type": "inline_equation" }, { "bbox": [ 223, 473, 227, 489 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 227, 475, 254, 487 ], "score": 0.85, "content": "\\lambda _ { \\mathrm { F G \\mathrm { - } c l s } }", "type": "inline_equation" }, { "bbox": [ 254, 473, 272, 489 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 272, 475, 299, 487 ], "score": 0.91, "content": "\\lambda _ { \\mathrm { B G \\mathrm { { - } c l s } } }", "type": "inline_equation" }, { "bbox": [ 300, 473, 458, 489 ], "score": 1.0, "content": "are the hyperparameters for each term.", "type": "text" } ], "index": 28 } ], "index": 28, "bbox_fs": [ 105, 473, 458, 489 ] }, { "type": "title", "bbox": [ 107, 502, 200, 515 ], "lines": [ { "bbox": [ 105, 502, 201, 516 ], "spans": [ { "bbox": [ 105, 502, 201, 516 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 29 } ], "index": 29 }, { "type": "text", "bbox": [ 108, 527, 504, 560 ], "lines": [ { "bbox": [ 105, 525, 505, 540 ], "spans": [ { "bbox": [ 105, 525, 505, 540 ], "score": 1.0, "content": "We evaluated the images generated by DT-GAN through a series of experiments both quantitatively", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 538, 505, 550 ], "spans": [ { "bbox": [ 106, 538, 505, 550 ], "score": 1.0, "content": "and qualitatively. Finally, we demonstrate the benefits of our generated images when being used as", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 549, 368, 561 ], "spans": [ { "bbox": [ 106, 549, 368, 561 ], "score": 1.0, "content": "data augmentation for a defect classification task on limited data.", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 105, 525, 505, 561 ] }, { "type": "text", "bbox": [ 106, 565, 505, 665 ], "lines": [ { "bbox": [ 106, 566, 505, 577 ], "spans": [ { "bbox": [ 106, 566, 505, 577 ], "score": 1.0, "content": "Dataset. All experiments were performed on a real industrial dataset: a Surface Defect Inspection", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 577, 504, 589 ], "spans": [ { "bbox": [ 106, 577, 504, 589 ], "score": 1.0, "content": "(SDI) dataset that contains three different kinds of products from production lines and samples from", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 588, 505, 600 ], "spans": [ { "bbox": [ 106, 588, 505, 600 ], "score": 1.0, "content": "each product are classified into three mutually exclusive classes: Normal, Scratch and Spot.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 598, 505, 612 ], "spans": [ { "bbox": [ 105, 598, 505, 612 ], "score": 1.0, "content": "All of the images are grayscale. Detailed statistics of the dataset are summarized in Appendix", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 609, 506, 622 ], "spans": [ { "bbox": [ 105, 609, 506, 622 ], "score": 1.0, "content": "A. Note that only the training set was used in GAN training, the test set was left untouched for", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 621, 504, 634 ], "spans": [ { "bbox": [ 106, 621, 459, 634 ], "score": 1.0, "content": "final evaluation in classifier training. For a fair comparison, all images were resized to", "type": "text" }, { "bbox": [ 460, 621, 504, 631 ], "score": 0.89, "content": "1 2 8 \\times 1 2 8", "type": "inline_equation" } ], "index": 38 }, { "bbox": [ 106, 632, 505, 644 ], "spans": [ { "bbox": [ 106, 632, 505, 644 ], "score": 1.0, "content": "resolution for both GAN training and classifier training, which was also the highest resolution used", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 642, 505, 656 ], "spans": [ { "bbox": [ 105, 642, 505, 656 ], "score": 1.0, "content": "in the baselines for image generation. For comparison, we also conducted experiments on the widely", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 653, 439, 666 ], "spans": [ { "bbox": [ 105, 653, 439, 666 ], "score": 1.0, "content": "used MVTec Anomaly Detection dataset (Bergmann et al., 2019) in Appendix E.4.", "type": "text" } ], "index": 41 } ], "index": 37, "bbox_fs": [ 105, 566, 506, 666 ] }, { "type": "title", "bbox": [ 108, 678, 225, 689 ], "lines": [ { "bbox": [ 105, 677, 227, 691 ], "spans": [ { "bbox": [ 105, 677, 227, 691 ], "score": 1.0, "content": "4.1 DEFECT GENERATION", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "Baselines. As discussed in Section 3, DT-GAN can either use the mapping network to randomly", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "generate styles and defects, or it can use the style-content encoder to extract both from reference", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 720, 454, 734 ], "spans": [ { "bbox": [ 105, 720, 454, 734 ], "score": 1.0, "content": "images. We refer to these cases as ‘latent-guided’ and ‘reference-guided’, respectively.", "type": "text" } ], "index": 45 } ], "index": 44, "bbox_fs": [ 105, 698, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 182 ], "lines": [ { "bbox": [ 105, 82, 506, 95 ], "spans": [ { "bbox": [ 105, 82, 506, 95 ], "score": 1.0, "content": "Since the two ways of guidance are fundamentally different, we evaluated them against two sets", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 506, 107 ], "spans": [ { "bbox": [ 105, 93, 506, 107 ], "score": 1.0, "content": "of baselines: Our reference-guided image generation was compared to Mokady et al. (2020) and", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 105, 504, 117 ], "spans": [ { "bbox": [ 106, 105, 504, 117 ], "score": 1.0, "content": "StarGAN v2, because both of them can perform a reference-guided translation. Note that Mokady", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 506, 128 ], "spans": [ { "bbox": [ 105, 115, 506, 128 ], "score": 1.0, "content": "et al. (2020) can only translate between two domains while StarGAN v2 and DT-GAN can achieve", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 126, 506, 140 ], "spans": [ { "bbox": [ 105, 126, 506, 140 ], "score": 1.0, "content": "multi-domain translation within a single model. Images generated through the latent-guided part", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 136, 505, 150 ], "spans": [ { "bbox": [ 105, 136, 505, 150 ], "score": 1.0, "content": "of DT-GAN were compared to state-of-the-art GANs in image synthesis: BigGAN (Brock et al.,", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 162 ], "spans": [ { "bbox": [ 105, 147, 505, 162 ], "score": 1.0, "content": "2019) and StyleGAN v2 (Karras et al., 2020b). We set BigGAN to condition on defect types during", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 505, 172 ], "spans": [ { "bbox": [ 105, 159, 505, 172 ], "score": 1.0, "content": "training while StyleGAN v2 was trained unconditionally. All baselines were trained from scratch", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 171, 340, 182 ], "spans": [ { "bbox": [ 106, 171, 340, 182 ], "score": 1.0, "content": "with the public implementations provided by the authors1.", "type": "text" } ], "index": 8 } ], "index": 4 }, { "type": "title", "bbox": [ 108, 195, 262, 206 ], "lines": [ { "bbox": [ 106, 194, 264, 208 ], "spans": [ { "bbox": [ 106, 194, 264, 208 ], "score": 1.0, "content": "4.1.1 QUANTITATIVE EVALUATION", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 107, 214, 505, 259 ], "lines": [ { "bbox": [ 105, 214, 506, 227 ], "spans": [ { "bbox": [ 105, 214, 506, 227 ], "score": 1.0, "content": "Metrics. We employed the commonly used frechet inception distance (FID) (Heusel et al., 2017) to", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 226, 506, 239 ], "spans": [ { "bbox": [ 105, 226, 506, 239 ], "score": 1.0, "content": "evaluate both the visual quality and the diversity of the generated images. We also report the kernel", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 237, 506, 249 ], "spans": [ { "bbox": [ 106, 237, 506, 249 ], "score": 1.0, "content": "inception distance (KID) (Binkowski et al., 2018) which is a more stable metric for small sets of", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 248, 448, 261 ], "spans": [ { "bbox": [ 105, 248, 448, 261 ], "score": 1.0, "content": "images like our SDI dataset. Lower FID and KID scores indicate better performance.", "type": "text" } ], "index": 13 } ], "index": 11.5 }, { "type": "text", "bbox": [ 107, 264, 505, 331 ], "lines": [ { "bbox": [ 106, 265, 505, 276 ], "spans": [ { "bbox": [ 106, 265, 505, 276 ], "score": 1.0, "content": "Both scores are shown in Table 1. We observe that methods like BigGAN and StyleGAN v2, which", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 276, 505, 288 ], "spans": [ { "bbox": [ 105, 276, 505, 288 ], "score": 1.0, "content": "perform defect synthesis purely based on latent codes, generally provide unsatisfactory results on", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 287, 505, 299 ], "spans": [ { "bbox": [ 106, 287, 505, 299 ], "score": 1.0, "content": "the SDI dataset, presumably due to the small number of defective samples that were available.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 297, 505, 310 ], "spans": [ { "bbox": [ 105, 297, 505, 310 ], "score": 1.0, "content": "These methods then struggle to capture the complex and irregular patterns of defects. We also", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 308, 505, 321 ], "spans": [ { "bbox": [ 105, 308, 505, 321 ], "score": 1.0, "content": "experimented with augmentation methods for GAN training (Karras et al., 2020a; Zhao et al., 2020)", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 318, 502, 332 ], "spans": [ { "bbox": [ 105, 318, 502, 332 ], "score": 1.0, "content": "but did not find a consistent improvement (see Appendix E.2). We thus only report the best scores.", "type": "text" } ], "index": 19 } ], "index": 16.5 }, { "type": "text", "bbox": [ 107, 336, 505, 414 ], "lines": [ { "bbox": [ 105, 335, 506, 349 ], "spans": [ { "bbox": [ 105, 335, 506, 349 ], "score": 1.0, "content": "Reference-guided synthesis methods like Mokady et al. (2020) and StarGAN v2 seem to generate", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 347, 505, 360 ], "spans": [ { "bbox": [ 105, 347, 505, 360 ], "score": 1.0, "content": "more realistic images. The scores of StarGAN v2 on a single product are omitted here because", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 358, 506, 371 ], "spans": [ { "bbox": [ 105, 358, 506, 371 ], "score": 1.0, "content": "generating images with specified background is not possible due to its network design—the product", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 370, 505, 381 ], "spans": [ { "bbox": [ 105, 370, 505, 381 ], "score": 1.0, "content": "type changes in output images, which we refer to as ‘identity-shift’. As seen in Table 1, our method", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 380, 505, 392 ], "spans": [ { "bbox": [ 105, 380, 505, 392 ], "score": 1.0, "content": "achieves better scores in all cases. We believe this is due to the fact that our method allows free", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 390, 505, 404 ], "spans": [ { "bbox": [ 105, 390, 505, 404 ], "score": 1.0, "content": "combination of foreground defects and backgrounds, making the generated images more diverse", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 403, 293, 415 ], "spans": [ { "bbox": [ 105, 403, 293, 415 ], "score": 1.0, "content": "even with a small number of training samples.", "type": "text" } ], "index": 26 } ], "index": 23 }, { "type": "table", "bbox": [ 108, 464, 499, 554 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 430, 502, 464 ], "group_id": 0, "lines": [ { "bbox": [ 105, 430, 504, 443 ], "spans": [ { "bbox": [ 105, 430, 504, 443 ], "score": 1.0, "content": "Table 1: Quantitative comparison of DT-GAN with baseline image synthesis methods using FID", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 441, 504, 453 ], "spans": [ { "bbox": [ 106, 441, 504, 453 ], "score": 1.0, "content": "and KID. Note that the reported values are not comparable between columns, because they were", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 452, 252, 465 ], "spans": [ { "bbox": [ 105, 452, 252, 465 ], "score": 1.0, "content": "calculated on different training sets.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "table_body", "bbox": [ 108, 464, 499, 554 ], "group_id": 0, "lines": [ { "bbox": [ 108, 464, 499, 554 ], "spans": [ { "bbox": [ 108, 464, 499, 554 ], "score": 0.979, "html": "
MethodFID↓KID↓
ABCAllABCAll
Mokady (2020)68.6966.9036.2158.630.0500.0360.0300.036
StarGAN v21137.70-110.013
StyleGAN v290.1052.95138.0935.340.0720.0270.1860.013
BigGAN + DiffAug218.74134.41270.89155.880.2200.1210.3780.099
Ours58.4336.4422.6829.730.0250.0130.0120.009
", "type": "table", "image_path": "7998b6b2f82acd4e8a5d2d8cfa152b29cacad12cfc7d986a268c2db781bbb119.jpg" } ] } ], "index": 31, "virtual_lines": [ { "bbox": [ 108, 464, 499, 494.0 ], "spans": [], "index": 30 }, { "bbox": [ 108, 494.0, 499, 524.0 ], "spans": [], "index": 31 }, { "bbox": [ 108, 524.0, 499, 554.0 ], "spans": [], "index": 32 } ] } ], "index": 29.5 }, { "type": "title", "bbox": [ 108, 570, 256, 582 ], "lines": [ { "bbox": [ 106, 570, 258, 583 ], "spans": [ { "bbox": [ 106, 570, 258, 583 ], "score": 1.0, "content": "4.1.2 QUALITATIVE EVALUATION", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 107, 589, 505, 711 ], "lines": [ { "bbox": [ 106, 590, 505, 602 ], "spans": [ { "bbox": [ 106, 590, 505, 602 ], "score": 1.0, "content": "We present a qualitative comparison with the baseline methods in latent-guided image synthesis", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 601, 505, 613 ], "spans": [ { "bbox": [ 105, 601, 505, 613 ], "score": 1.0, "content": "in Figure 3. To make a fair comparison, we trained StyleGAN v2 and BigGAN on each product", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 611, 506, 625 ], "spans": [ { "bbox": [ 105, 611, 506, 625 ], "score": 1.0, "content": "separately to have control on background products. Note however, that images from DT-GAN were", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 622, 505, 636 ], "spans": [ { "bbox": [ 105, 622, 505, 636 ], "score": 1.0, "content": "always obtained from a single model. We can see that some generated samples from StyleGAN", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 633, 505, 646 ], "spans": [ { "bbox": [ 106, 633, 505, 646 ], "score": 1.0, "content": "v2 do not contain clear defects, and samples from BigGAN present abnormal grid patterns. Both", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 644, 505, 657 ], "spans": [ { "bbox": [ 105, 644, 505, 657 ], "score": 1.0, "content": "methods do not take images as inputs but generate synthetic images according to a given latent code", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 654, 505, 668 ], "spans": [ { "bbox": [ 106, 654, 505, 668 ], "score": 1.0, "content": "which contains information for both FG and BG. This conditioning leads to limited diversity in", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 667, 506, 679 ], "spans": [ { "bbox": [ 105, 667, 506, 679 ], "score": 1.0, "content": "the output images. On the other hand, StarGAN v2 performs translation based on input images but", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 677, 505, 690 ], "spans": [ { "bbox": [ 105, 677, 505, 690 ], "score": 1.0, "content": "suffers from the same entanglement issue. Thus, it fails to preserve the background, which results in", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 505, 701 ], "score": 1.0, "content": "artifacts or identity-shift in its outputs. Our network architecture that disentangles foreground and", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 699, 425, 713 ], "spans": [ { "bbox": [ 106, 699, 425, 713 ], "score": 1.0, "content": "background seems to mitigate these issues. See Appendix E.4 for more images.", "type": "text" } ], "index": 44 } ], "index": 39 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 119, 721, 381, 731 ], "lines": [ { "bbox": [ 119, 720, 381, 733 ], "spans": [ { "bbox": [ 119, 720, 381, 733 ], "score": 1.0, "content": "1We could not obtain the code of Defect-GAN to reproduce their results.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 309, 39 ], "spans": [ { "bbox": [ 106, 25, 309, 39 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 182 ], "lines": [ { "bbox": [ 105, 82, 506, 95 ], "spans": [ { "bbox": [ 105, 82, 506, 95 ], "score": 1.0, "content": "Since the two ways of guidance are fundamentally different, we evaluated them against two sets", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 506, 107 ], "spans": [ { "bbox": [ 105, 93, 506, 107 ], "score": 1.0, "content": "of baselines: Our reference-guided image generation was compared to Mokady et al. (2020) and", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 105, 504, 117 ], "spans": [ { "bbox": [ 106, 105, 504, 117 ], "score": 1.0, "content": "StarGAN v2, because both of them can perform a reference-guided translation. Note that Mokady", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 506, 128 ], "spans": [ { "bbox": [ 105, 115, 506, 128 ], "score": 1.0, "content": "et al. (2020) can only translate between two domains while StarGAN v2 and DT-GAN can achieve", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 126, 506, 140 ], "spans": [ { "bbox": [ 105, 126, 506, 140 ], "score": 1.0, "content": "multi-domain translation within a single model. Images generated through the latent-guided part", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 136, 505, 150 ], "spans": [ { "bbox": [ 105, 136, 505, 150 ], "score": 1.0, "content": "of DT-GAN were compared to state-of-the-art GANs in image synthesis: BigGAN (Brock et al.,", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 162 ], "spans": [ { "bbox": [ 105, 147, 505, 162 ], "score": 1.0, "content": "2019) and StyleGAN v2 (Karras et al., 2020b). We set BigGAN to condition on defect types during", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 505, 172 ], "spans": [ { "bbox": [ 105, 159, 505, 172 ], "score": 1.0, "content": "training while StyleGAN v2 was trained unconditionally. All baselines were trained from scratch", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 171, 340, 182 ], "spans": [ { "bbox": [ 106, 171, 340, 182 ], "score": 1.0, "content": "with the public implementations provided by the authors1.", "type": "text" } ], "index": 8 } ], "index": 4, "bbox_fs": [ 105, 82, 506, 182 ] }, { "type": "title", "bbox": [ 108, 195, 262, 206 ], "lines": [ { "bbox": [ 106, 194, 264, 208 ], "spans": [ { "bbox": [ 106, 194, 264, 208 ], "score": 1.0, "content": "4.1.1 QUANTITATIVE EVALUATION", "type": "text" } ], "index": 9 } ], "index": 9 }, { "type": "text", "bbox": [ 107, 214, 505, 259 ], "lines": [ { "bbox": [ 105, 214, 506, 227 ], "spans": [ { "bbox": [ 105, 214, 506, 227 ], "score": 1.0, "content": "Metrics. We employed the commonly used frechet inception distance (FID) (Heusel et al., 2017) to", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 226, 506, 239 ], "spans": [ { "bbox": [ 105, 226, 506, 239 ], "score": 1.0, "content": "evaluate both the visual quality and the diversity of the generated images. We also report the kernel", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 237, 506, 249 ], "spans": [ { "bbox": [ 106, 237, 506, 249 ], "score": 1.0, "content": "inception distance (KID) (Binkowski et al., 2018) which is a more stable metric for small sets of", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 248, 448, 261 ], "spans": [ { "bbox": [ 105, 248, 448, 261 ], "score": 1.0, "content": "images like our SDI dataset. Lower FID and KID scores indicate better performance.", "type": "text" } ], "index": 13 } ], "index": 11.5, "bbox_fs": [ 105, 214, 506, 261 ] }, { "type": "text", "bbox": [ 107, 264, 505, 331 ], "lines": [ { "bbox": [ 106, 265, 505, 276 ], "spans": [ { "bbox": [ 106, 265, 505, 276 ], "score": 1.0, "content": "Both scores are shown in Table 1. We observe that methods like BigGAN and StyleGAN v2, which", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 276, 505, 288 ], "spans": [ { "bbox": [ 105, 276, 505, 288 ], "score": 1.0, "content": "perform defect synthesis purely based on latent codes, generally provide unsatisfactory results on", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 287, 505, 299 ], "spans": [ { "bbox": [ 106, 287, 505, 299 ], "score": 1.0, "content": "the SDI dataset, presumably due to the small number of defective samples that were available.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 297, 505, 310 ], "spans": [ { "bbox": [ 105, 297, 505, 310 ], "score": 1.0, "content": "These methods then struggle to capture the complex and irregular patterns of defects. We also", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 308, 505, 321 ], "spans": [ { "bbox": [ 105, 308, 505, 321 ], "score": 1.0, "content": "experimented with augmentation methods for GAN training (Karras et al., 2020a; Zhao et al., 2020)", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 318, 502, 332 ], "spans": [ { "bbox": [ 105, 318, 502, 332 ], "score": 1.0, "content": "but did not find a consistent improvement (see Appendix E.2). We thus only report the best scores.", "type": "text" } ], "index": 19 } ], "index": 16.5, "bbox_fs": [ 105, 265, 505, 332 ] }, { "type": "text", "bbox": [ 107, 336, 505, 414 ], "lines": [ { "bbox": [ 105, 335, 506, 349 ], "spans": [ { "bbox": [ 105, 335, 506, 349 ], "score": 1.0, "content": "Reference-guided synthesis methods like Mokady et al. (2020) and StarGAN v2 seem to generate", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 347, 505, 360 ], "spans": [ { "bbox": [ 105, 347, 505, 360 ], "score": 1.0, "content": "more realistic images. The scores of StarGAN v2 on a single product are omitted here because", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 358, 506, 371 ], "spans": [ { "bbox": [ 105, 358, 506, 371 ], "score": 1.0, "content": "generating images with specified background is not possible due to its network design—the product", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 370, 505, 381 ], "spans": [ { "bbox": [ 105, 370, 505, 381 ], "score": 1.0, "content": "type changes in output images, which we refer to as ‘identity-shift’. As seen in Table 1, our method", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 380, 505, 392 ], "spans": [ { "bbox": [ 105, 380, 505, 392 ], "score": 1.0, "content": "achieves better scores in all cases. We believe this is due to the fact that our method allows free", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 390, 505, 404 ], "spans": [ { "bbox": [ 105, 390, 505, 404 ], "score": 1.0, "content": "combination of foreground defects and backgrounds, making the generated images more diverse", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 403, 293, 415 ], "spans": [ { "bbox": [ 105, 403, 293, 415 ], "score": 1.0, "content": "even with a small number of training samples.", "type": "text" } ], "index": 26 } ], "index": 23, "bbox_fs": [ 105, 335, 506, 415 ] }, { "type": "table", "bbox": [ 108, 464, 499, 554 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 430, 502, 464 ], "group_id": 0, "lines": [ { "bbox": [ 105, 430, 504, 443 ], "spans": [ { "bbox": [ 105, 430, 504, 443 ], "score": 1.0, "content": "Table 1: Quantitative comparison of DT-GAN with baseline image synthesis methods using FID", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 441, 504, 453 ], "spans": [ { "bbox": [ 106, 441, 504, 453 ], "score": 1.0, "content": "and KID. Note that the reported values are not comparable between columns, because they were", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 452, 252, 465 ], "spans": [ { "bbox": [ 105, 452, 252, 465 ], "score": 1.0, "content": "calculated on different training sets.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "table_body", "bbox": [ 108, 464, 499, 554 ], "group_id": 0, "lines": [ { "bbox": [ 108, 464, 499, 554 ], "spans": [ { "bbox": [ 108, 464, 499, 554 ], "score": 0.979, "html": "
MethodFID↓KID↓
ABCAllABCAll
Mokady (2020)68.6966.9036.2158.630.0500.0360.0300.036
StarGAN v21137.70-110.013
StyleGAN v290.1052.95138.0935.340.0720.0270.1860.013
BigGAN + DiffAug218.74134.41270.89155.880.2200.1210.3780.099
Ours58.4336.4422.6829.730.0250.0130.0120.009
", "type": "table", "image_path": "7998b6b2f82acd4e8a5d2d8cfa152b29cacad12cfc7d986a268c2db781bbb119.jpg" } ] } ], "index": 31, "virtual_lines": [ { "bbox": [ 108, 464, 499, 494.0 ], "spans": [], "index": 30 }, { "bbox": [ 108, 494.0, 499, 524.0 ], "spans": [], "index": 31 }, { "bbox": [ 108, 524.0, 499, 554.0 ], "spans": [], "index": 32 } ] } ], "index": 29.5 }, { "type": "title", "bbox": [ 108, 570, 256, 582 ], "lines": [ { "bbox": [ 106, 570, 258, 583 ], "spans": [ { "bbox": [ 106, 570, 258, 583 ], "score": 1.0, "content": "4.1.2 QUALITATIVE EVALUATION", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 107, 589, 505, 711 ], "lines": [ { "bbox": [ 106, 590, 505, 602 ], "spans": [ { "bbox": [ 106, 590, 505, 602 ], "score": 1.0, "content": "We present a qualitative comparison with the baseline methods in latent-guided image synthesis", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 601, 505, 613 ], "spans": [ { "bbox": [ 105, 601, 505, 613 ], "score": 1.0, "content": "in Figure 3. To make a fair comparison, we trained StyleGAN v2 and BigGAN on each product", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 611, 506, 625 ], "spans": [ { "bbox": [ 105, 611, 506, 625 ], "score": 1.0, "content": "separately to have control on background products. Note however, that images from DT-GAN were", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 622, 505, 636 ], "spans": [ { "bbox": [ 105, 622, 505, 636 ], "score": 1.0, "content": "always obtained from a single model. We can see that some generated samples from StyleGAN", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 633, 505, 646 ], "spans": [ { "bbox": [ 106, 633, 505, 646 ], "score": 1.0, "content": "v2 do not contain clear defects, and samples from BigGAN present abnormal grid patterns. Both", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 644, 505, 657 ], "spans": [ { "bbox": [ 105, 644, 505, 657 ], "score": 1.0, "content": "methods do not take images as inputs but generate synthetic images according to a given latent code", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 654, 505, 668 ], "spans": [ { "bbox": [ 106, 654, 505, 668 ], "score": 1.0, "content": "which contains information for both FG and BG. This conditioning leads to limited diversity in", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 667, 506, 679 ], "spans": [ { "bbox": [ 105, 667, 506, 679 ], "score": 1.0, "content": "the output images. On the other hand, StarGAN v2 performs translation based on input images but", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 677, 505, 690 ], "spans": [ { "bbox": [ 105, 677, 505, 690 ], "score": 1.0, "content": "suffers from the same entanglement issue. Thus, it fails to preserve the background, which results in", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 505, 701 ], "score": 1.0, "content": "artifacts or identity-shift in its outputs. Our network architecture that disentangles foreground and", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 699, 425, 713 ], "spans": [ { "bbox": [ 106, 699, 425, 713 ], "score": 1.0, "content": "background seems to mitigate these issues. See Appendix E.4 for more images.", "type": "text" } ], "index": 44 } ], "index": 39, "bbox_fs": [ 105, 590, 506, 713 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 108, 82, 505, 212 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 82, 505, 212 ], "group_id": 0, "lines": [ { "bbox": [ 108, 82, 505, 212 ], "spans": [ { "bbox": [ 108, 82, 505, 212 ], "score": 0.962, "type": "image", "image_path": "e8844ad801507fcb58625110b5ddd1e08b278ccc393d4b6551f6e75e4f93f6ca.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 108, 82, 505, 125.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 108, 125.33333333333334, 505, 168.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 108, 168.66666666666669, 505, 212.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 215, 506, 248 ], "group_id": 0, "lines": [ { "bbox": [ 106, 215, 505, 227 ], "spans": [ { "bbox": [ 106, 215, 505, 227 ], "score": 1.0, "content": "Figure 3: Qualitative comparison of latent-guided image synthesis results. In each subfigure: on", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 226, 505, 238 ], "spans": [ { "bbox": [ 106, 226, 505, 238 ], "score": 1.0, "content": "the left, defective images are fully generated from random noise. On the right, random defects are", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 236, 500, 250 ], "spans": [ { "bbox": [ 105, 236, 500, 250 ], "score": 1.0, "content": "synthesized onto given normal samples. Note that BigGAN* denotes it was trained with DiffAug.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 112, 253, 500, 384 ], "blocks": [ { "type": "image_body", "bbox": [ 112, 253, 500, 384 ], "group_id": 1, "lines": [ { "bbox": [ 112, 253, 500, 384 ], "spans": [ { "bbox": [ 112, 253, 500, 384 ], "score": 0.945, "type": "image", "image_path": "c8d4f138dd7a663cb4fbcdd6d95ca7b5d88bc4505014c289d8d78e678a6ef857.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 112, 253, 500, 296.6666666666667 ], "spans": [], "index": 6 }, { "bbox": [ 112, 296.6666666666667, 500, 340.33333333333337 ], "spans": [], "index": 7 }, { "bbox": [ 112, 340.33333333333337, 500, 384.00000000000006 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 107, 386, 505, 419 ], "group_id": 1, "lines": [ { "bbox": [ 105, 385, 505, 398 ], "spans": [ { "bbox": [ 105, 385, 505, 398 ], "score": 1.0, "content": "Figure 4: Qualitative comparison of reference-guided image synthesis results on the SDI", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 394, 506, 411 ], "spans": [ { "bbox": [ 105, 394, 506, 411 ], "score": 1.0, "content": "dataset. Each method transforms the given source images into target foreground domains (e.g.,", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 407, 427, 421 ], "spans": [ { "bbox": [ 105, 407, 427, 421 ], "score": 1.0, "content": "Scratches) with the styles and contents extracted from the reference images.", "type": "text" } ], "index": 11 } ], "index": 10 } ], "index": 8.5 }, { "type": "text", "bbox": [ 107, 444, 505, 478 ], "lines": [ { "bbox": [ 106, 444, 505, 457 ], "spans": [ { "bbox": [ 106, 444, 505, 457 ], "score": 1.0, "content": "Also for reference-guided image synthesis, where we used different background and foreground", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "reference images as illustrated in Figure 4, only our method produces high quality images with", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 466, 477, 479 ], "spans": [ { "bbox": [ 105, 466, 477, 479 ], "score": 1.0, "content": "preserved background from the source and transferred foreground defect from the reference.", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 483, 505, 516 ], "lines": [ { "bbox": [ 106, 483, 505, 496 ], "spans": [ { "bbox": [ 106, 483, 505, 496 ], "score": 1.0, "content": "Ablation study. We visually demonstrate the effect of each component we added to DT-GAN", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 494, 506, 508 ], "spans": [ { "bbox": [ 105, 494, 506, 508 ], "score": 1.0, "content": "compared to StarGAN v2 in Figure 5, using the examples of both latent- and reference-guided image", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 505, 505, 519 ], "spans": [ { "bbox": [ 105, 505, 505, 519 ], "score": 1.0, "content": "synthesis from Normal to Scratches. The quantitative evaluation can be found in Appendix E.3.", "type": "text" } ], "index": 17 } ], "index": 16 }, { "type": "text", "bbox": [ 106, 522, 505, 611 ], "lines": [ { "bbox": [ 106, 522, 505, 535 ], "spans": [ { "bbox": [ 106, 522, 505, 535 ], "score": 1.0, "content": "Column (a) corresponds to StarGAN v2 and highlights the drawback of entangled FG/BG again", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 533, 505, 546 ], "spans": [ { "bbox": [ 106, 533, 505, 546 ], "score": 1.0, "content": "(i.e. the identity-shift in the background). We first tackle this problem by modeling the style code", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 544, 506, 558 ], "spans": [ { "bbox": [ 105, 544, 506, 558 ], "score": 1.0, "content": "and foreground content explicitly and feeding them separately to the generator. This leads to a better", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 555, 506, 568 ], "spans": [ { "bbox": [ 104, 555, 506, 568 ], "score": 1.0, "content": "preservation of the background structure in column (b) for the reference-guided subnetwork, but not", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 565, 505, 578 ], "spans": [ { "bbox": [ 105, 565, 505, 578 ], "score": 1.0, "content": "for the latent-guided synthesis on the bottom of Figure 5. Thus, we add a foreground classifier in", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 577, 505, 589 ], "spans": [ { "bbox": [ 106, 577, 505, 589 ], "score": 1.0, "content": "the discriminator in (c) to ensure the output image contains the desired foreground content (scratch).", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 588, 505, 600 ], "spans": [ { "bbox": [ 106, 588, 505, 600 ], "score": 1.0, "content": "Similarly, we introduce a background classifier to the discriminator in column (d). Note that the", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 599, 441, 612 ], "spans": [ { "bbox": [ 106, 599, 441, 612 ], "score": 1.0, "content": "additional product type labels can be acquired automatically from production lines.", "type": "text" } ], "index": 25 } ], "index": 21.5 }, { "type": "text", "bbox": [ 107, 616, 505, 693 ], "lines": [ { "bbox": [ 105, 616, 505, 628 ], "spans": [ { "bbox": [ 105, 616, 505, 628 ], "score": 1.0, "content": "For column (e), we add the separate decoders for foreground and background in the generator which", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 626, 505, 640 ], "spans": [ { "bbox": [ 105, 626, 505, 640 ], "score": 1.0, "content": "are fused only in the end. This enhances the preservation of background characteristics like lighting", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 638, 505, 650 ], "spans": [ { "bbox": [ 105, 638, 505, 650 ], "score": 1.0, "content": "even more. Imposing an additional penalty for foreground content extracted from a normal sam-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "ple as described in Section 3.1 leads to another visual improvement of the foreground edges for", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 659, 505, 673 ], "spans": [ { "bbox": [ 105, 659, 505, 673 ], "score": 1.0, "content": "reference-guided synthesis in column (f). Finally, inspired by StyleGAN, we incorporate adaptive", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 671, 505, 683 ], "spans": [ { "bbox": [ 105, 671, 505, 683 ], "score": 1.0, "content": "noise injection to the mapping network, which significantly boosts the performance of our latent-", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 682, 300, 694 ], "spans": [ { "bbox": [ 106, 682, 300, 694 ], "score": 1.0, "content": "guided image synthesis as shown in column (g).", "type": "text" } ], "index": 32 } ], "index": 29 }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 696, 506, 713 ], "spans": [ { "bbox": [ 105, 696, 506, 713 ], "score": 1.0, "content": "Styling. We visually demonstrate the effect of style codes in our method by randomly sampling", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "those and combining them with fixed reference background and foreground images in Figure 6,", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 721, 378, 733 ], "spans": [ { "bbox": [ 106, 721, 378, 733 ], "score": 1.0, "content": "where a variety of artistic styles can be seen on the output columns.", "type": "text" } ], "index": 35 } ], "index": 34 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 309, 39 ], "spans": [ { "bbox": [ 106, 25, 309, 39 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 303, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 750, 309, 762 ], "spans": [ { "bbox": [ 302, 750, 309, 762 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 108, 82, 505, 212 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 82, 505, 212 ], "group_id": 0, "lines": [ { "bbox": [ 108, 82, 505, 212 ], "spans": [ { "bbox": [ 108, 82, 505, 212 ], "score": 0.962, "type": "image", "image_path": "e8844ad801507fcb58625110b5ddd1e08b278ccc393d4b6551f6e75e4f93f6ca.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 108, 82, 505, 125.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 108, 125.33333333333334, 505, 168.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 108, 168.66666666666669, 505, 212.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 215, 506, 248 ], "group_id": 0, "lines": [ { "bbox": [ 106, 215, 505, 227 ], "spans": [ { "bbox": [ 106, 215, 505, 227 ], "score": 1.0, "content": "Figure 3: Qualitative comparison of latent-guided image synthesis results. In each subfigure: on", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 226, 505, 238 ], "spans": [ { "bbox": [ 106, 226, 505, 238 ], "score": 1.0, "content": "the left, defective images are fully generated from random noise. On the right, random defects are", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 236, 500, 250 ], "spans": [ { "bbox": [ 105, 236, 500, 250 ], "score": 1.0, "content": "synthesized onto given normal samples. Note that BigGAN* denotes it was trained with DiffAug.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 112, 253, 500, 384 ], "blocks": [ { "type": "image_body", "bbox": [ 112, 253, 500, 384 ], "group_id": 1, "lines": [ { "bbox": [ 112, 253, 500, 384 ], "spans": [ { "bbox": [ 112, 253, 500, 384 ], "score": 0.945, "type": "image", "image_path": "c8d4f138dd7a663cb4fbcdd6d95ca7b5d88bc4505014c289d8d78e678a6ef857.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 112, 253, 500, 296.6666666666667 ], "spans": [], "index": 6 }, { "bbox": [ 112, 296.6666666666667, 500, 340.33333333333337 ], "spans": [], "index": 7 }, { "bbox": [ 112, 340.33333333333337, 500, 384.00000000000006 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 107, 386, 505, 419 ], "group_id": 1, "lines": [ { "bbox": [ 105, 385, 505, 398 ], "spans": [ { "bbox": [ 105, 385, 505, 398 ], "score": 1.0, "content": "Figure 4: Qualitative comparison of reference-guided image synthesis results on the SDI", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 394, 506, 411 ], "spans": [ { "bbox": [ 105, 394, 506, 411 ], "score": 1.0, "content": "dataset. Each method transforms the given source images into target foreground domains (e.g.,", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 407, 427, 421 ], "spans": [ { "bbox": [ 105, 407, 427, 421 ], "score": 1.0, "content": "Scratches) with the styles and contents extracted from the reference images.", "type": "text" } ], "index": 11 } ], "index": 10 } ], "index": 8.5 }, { "type": "text", "bbox": [ 107, 444, 505, 478 ], "lines": [ { "bbox": [ 106, 444, 505, 457 ], "spans": [ { "bbox": [ 106, 444, 505, 457 ], "score": 1.0, "content": "Also for reference-guided image synthesis, where we used different background and foreground", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "reference images as illustrated in Figure 4, only our method produces high quality images with", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 466, 477, 479 ], "spans": [ { "bbox": [ 105, 466, 477, 479 ], "score": 1.0, "content": "preserved background from the source and transferred foreground defect from the reference.", "type": "text" } ], "index": 14 } ], "index": 13, "bbox_fs": [ 105, 444, 505, 479 ] }, { "type": "text", "bbox": [ 107, 483, 505, 516 ], "lines": [ { "bbox": [ 106, 483, 505, 496 ], "spans": [ { "bbox": [ 106, 483, 505, 496 ], "score": 1.0, "content": "Ablation study. We visually demonstrate the effect of each component we added to DT-GAN", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 494, 506, 508 ], "spans": [ { "bbox": [ 105, 494, 506, 508 ], "score": 1.0, "content": "compared to StarGAN v2 in Figure 5, using the examples of both latent- and reference-guided image", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 505, 505, 519 ], "spans": [ { "bbox": [ 105, 505, 505, 519 ], "score": 1.0, "content": "synthesis from Normal to Scratches. The quantitative evaluation can be found in Appendix E.3.", "type": "text" } ], "index": 17 } ], "index": 16, "bbox_fs": [ 105, 483, 506, 519 ] }, { "type": "text", "bbox": [ 106, 522, 505, 611 ], "lines": [ { "bbox": [ 106, 522, 505, 535 ], "spans": [ { "bbox": [ 106, 522, 505, 535 ], "score": 1.0, "content": "Column (a) corresponds to StarGAN v2 and highlights the drawback of entangled FG/BG again", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 533, 505, 546 ], "spans": [ { "bbox": [ 106, 533, 505, 546 ], "score": 1.0, "content": "(i.e. the identity-shift in the background). We first tackle this problem by modeling the style code", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 544, 506, 558 ], "spans": [ { "bbox": [ 105, 544, 506, 558 ], "score": 1.0, "content": "and foreground content explicitly and feeding them separately to the generator. This leads to a better", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 555, 506, 568 ], "spans": [ { "bbox": [ 104, 555, 506, 568 ], "score": 1.0, "content": "preservation of the background structure in column (b) for the reference-guided subnetwork, but not", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 565, 505, 578 ], "spans": [ { "bbox": [ 105, 565, 505, 578 ], "score": 1.0, "content": "for the latent-guided synthesis on the bottom of Figure 5. Thus, we add a foreground classifier in", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 577, 505, 589 ], "spans": [ { "bbox": [ 106, 577, 505, 589 ], "score": 1.0, "content": "the discriminator in (c) to ensure the output image contains the desired foreground content (scratch).", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 588, 505, 600 ], "spans": [ { "bbox": [ 106, 588, 505, 600 ], "score": 1.0, "content": "Similarly, we introduce a background classifier to the discriminator in column (d). Note that the", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 599, 441, 612 ], "spans": [ { "bbox": [ 106, 599, 441, 612 ], "score": 1.0, "content": "additional product type labels can be acquired automatically from production lines.", "type": "text" } ], "index": 25 } ], "index": 21.5, "bbox_fs": [ 104, 522, 506, 612 ] }, { "type": "text", "bbox": [ 107, 616, 505, 693 ], "lines": [ { "bbox": [ 105, 616, 505, 628 ], "spans": [ { "bbox": [ 105, 616, 505, 628 ], "score": 1.0, "content": "For column (e), we add the separate decoders for foreground and background in the generator which", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 626, 505, 640 ], "spans": [ { "bbox": [ 105, 626, 505, 640 ], "score": 1.0, "content": "are fused only in the end. This enhances the preservation of background characteristics like lighting", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 638, 505, 650 ], "spans": [ { "bbox": [ 105, 638, 505, 650 ], "score": 1.0, "content": "even more. Imposing an additional penalty for foreground content extracted from a normal sam-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 648, 505, 662 ], "spans": [ { "bbox": [ 105, 648, 505, 662 ], "score": 1.0, "content": "ple as described in Section 3.1 leads to another visual improvement of the foreground edges for", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 659, 505, 673 ], "spans": [ { "bbox": [ 105, 659, 505, 673 ], "score": 1.0, "content": "reference-guided synthesis in column (f). Finally, inspired by StyleGAN, we incorporate adaptive", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 671, 505, 683 ], "spans": [ { "bbox": [ 105, 671, 505, 683 ], "score": 1.0, "content": "noise injection to the mapping network, which significantly boosts the performance of our latent-", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 682, 300, 694 ], "spans": [ { "bbox": [ 106, 682, 300, 694 ], "score": 1.0, "content": "guided image synthesis as shown in column (g).", "type": "text" } ], "index": 32 } ], "index": 29, "bbox_fs": [ 105, 616, 505, 694 ] }, { "type": "text", "bbox": [ 108, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 696, 506, 713 ], "spans": [ { "bbox": [ 105, 696, 506, 713 ], "score": 1.0, "content": "Styling. We visually demonstrate the effect of style codes in our method by randomly sampling", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "those and combining them with fixed reference background and foreground images in Figure 6,", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 721, 378, 733 ], "spans": [ { "bbox": [ 106, 721, 378, 733 ], "score": 1.0, "content": "where a variety of artistic styles can be seen on the output columns.", "type": "text" } ], "index": 35 } ], "index": 34, "bbox_fs": [ 105, 696, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 81, 478, 246 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 81, 478, 246 ], "group_id": 0, "lines": [ { "bbox": [ 132, 81, 478, 246 ], "spans": [ { "bbox": [ 132, 81, 478, 246 ], "score": 0.968, "type": "image", "image_path": "91696fc85e55662030167a35ef4aae436fa0857c2b3a62e62f0e14e5967749b8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 81, 478, 136.0 ], "spans": [], "index": 0 }, { "bbox": [ 132, 136.0, 478, 191.0 ], "spans": [], "index": 1 }, { "bbox": [ 132, 191.0, 478, 246.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 258, 506, 292 ], "group_id": 0, "lines": [ { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 349, 271 ], "score": 1.0, "content": "Figure 5: Ablation study. (a) The baseline StarGAN v2. (b)", "type": "text" }, { "bbox": [ 349, 260, 358, 269 ], "score": 0.66, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 359, 258, 471, 271 ], "score": 1.0, "content": "Style-Content branches. (c)", "type": "text" }, { "bbox": [ 472, 260, 481, 269 ], "score": 0.72, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 481, 258, 505, 271 ], "score": 1.0, "content": "Fore-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 269, 506, 282 ], "spans": [ { "bbox": [ 105, 269, 190, 282 ], "score": 1.0, "content": "ground classifier. (d)", "type": "text" }, { "bbox": [ 190, 270, 199, 279 ], "score": 0.48, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 199, 269, 302, 282 ], "score": 1.0, "content": "Background classifier. (e)", "type": "text" }, { "bbox": [ 302, 271, 310, 279 ], "score": 0.64, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 311, 269, 506, 282 ], "score": 1.0, "content": "Separately decoding foreground and background", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 280, 502, 293 ], "spans": [ { "bbox": [ 105, 280, 117, 293 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 117, 281, 126, 290 ], "score": 0.53, "content": "G", "type": "inline_equation" }, { "bbox": [ 126, 280, 141, 293 ], "score": 1.0, "content": ". (f)", "type": "text" }, { "bbox": [ 142, 281, 150, 290 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 151, 280, 341, 293 ], "score": 1.0, "content": "Anchor foreground domain (e.g. Normal). (g)", "type": "text" }, { "bbox": [ 341, 281, 349, 290 ], "score": 0.59, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 350, 280, 502, 293 ], "score": 1.0, "content": "Noise injection in Mapping Network.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 112, 297, 499, 427 ], "blocks": [ { "type": "image_body", "bbox": [ 112, 297, 499, 427 ], "group_id": 1, "lines": [ { "bbox": [ 112, 297, 499, 427 ], "spans": [ { "bbox": [ 112, 297, 499, 427 ], "score": 0.954, "type": "image", "image_path": "10b4554eb02cdbab48f0a3b29e5fed2318aee10f044aacf5703500d8bb21eef0.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 112, 297, 499, 340.3333333333333 ], "spans": [], "index": 6 }, { "bbox": [ 112, 340.3333333333333, 499, 383.66666666666663 ], "spans": [], "index": 7 }, { "bbox": [ 112, 383.66666666666663, 499, 426.99999999999994 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 105, 429, 505, 452 ], "group_id": 1, "lines": [ { "bbox": [ 106, 429, 505, 442 ], "spans": [ { "bbox": [ 106, 429, 505, 442 ], "score": 1.0, "content": "Figure 6: Visual effect of randomly sampled style codes on fixed pairs of reference background", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 439, 280, 454 ], "spans": [ { "bbox": [ 105, 439, 280, 454 ], "score": 1.0, "content": "(Source) and foreground (Content) images.", "type": "text" } ], "index": 10 } ], "index": 9.5 } ], "index": 8.25 }, { "type": "title", "bbox": [ 108, 477, 289, 488 ], "lines": [ { "bbox": [ 106, 476, 291, 490 ], "spans": [ { "bbox": [ 106, 476, 291, 490 ], "score": 1.0, "content": "4.2 DT-GAN FOR DATA AUGMENTATION", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 500, 505, 533 ], "lines": [ { "bbox": [ 106, 500, 505, 512 ], "spans": [ { "bbox": [ 106, 500, 505, 512 ], "score": 1.0, "content": "We also evaluated our method as a data augmentation method for defect classification on the SDI", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 512, 504, 523 ], "spans": [ { "bbox": [ 106, 512, 504, 523 ], "score": 1.0, "content": "dataset. We defined one task ‘general’, where the classifier was trained on images from all products", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 522, 440, 534 ], "spans": [ { "bbox": [ 105, 522, 440, 534 ], "score": 1.0, "content": "at once, while task ‘single product’ only used the subset of images for one product.", "type": "text" } ], "index": 14 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 539, 505, 595 ], "lines": [ { "bbox": [ 105, 537, 505, 552 ], "spans": [ { "bbox": [ 105, 537, 505, 552 ], "score": 1.0, "content": "Besides, we incrementally varied the amount of real Normal data available for classifier training:", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 550, 505, 563 ], "spans": [ { "bbox": [ 105, 550, 505, 563 ], "score": 1.0, "content": "4500, 6600, 12000 and 18600. In the case of defective images, all of them were always used due to", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 560, 505, 572 ], "spans": [ { "bbox": [ 106, 560, 505, 572 ], "score": 1.0, "content": "the small amount unless otherwise specified. As backbone we used a ResNet-50 (He et al., 2016a)", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 570, 505, 586 ], "spans": [ { "bbox": [ 105, 570, 505, 586 ], "score": 1.0, "content": "with ImageNet pretrained weights. For experiments with synthetic data, we attached an auxiliary", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 583, 493, 595 ], "spans": [ { "bbox": [ 106, 583, 493, 595 ], "score": 1.0, "content": "domain classifier to the network through a Gradient Reversal Layer (Ganin & Lempitsky, 2015).", "type": "text" } ], "index": 19 } ], "index": 17 }, { "type": "table", "bbox": [ 163, 638, 444, 733 ], "blocks": [ { "type": "table_caption", "bbox": [ 105, 615, 504, 638 ], "group_id": 0, "lines": [ { "bbox": [ 106, 615, 505, 628 ], "spans": [ { "bbox": [ 106, 615, 505, 628 ], "score": 1.0, "content": "Table 2: Quantitative comparison of the baseline methods on defect classification task at the scale", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 626, 469, 639 ], "spans": [ { "bbox": [ 105, 626, 392, 639 ], "score": 1.0, "content": "of 12000 images/class. The reported values are the achieved error rates", "type": "text" }, { "bbox": [ 392, 627, 408, 637 ], "score": 0.78, "content": "( \\% )", "type": "inline_equation" }, { "bbox": [ 408, 626, 469, 639 ], "score": 1.0, "content": "over five runs.", "type": "text" } ], "index": 21 } ], "index": 20.5 }, { "type": "table_body", "bbox": [ 163, 638, 444, 733 ], "group_id": 0, "lines": [ { "bbox": [ 163, 638, 444, 733 ], "spans": [ { "bbox": [ 163, 638, 444, 733 ], "score": 0.899, "html": "
MethodResNet-50EfficientNet-b4
No-Aug21.64±1.2412.06±0.64
Trad-Aug12.58±0.819.33±0.73
Mokady (2020)11.11±1.1913.26±1.13
StarGAN v213.07±1.3012.25±0.79
StyleGAN v211.55±1.7911.68±0.76
BigGAN+DiffAug11.45±0.6112.06±0.50
Ours9.9±0.699.14±1.02
", "type": "table", "image_path": "928c69ed8d652dba082f99aeef3e7de87cc055491f1db2d2d419fae8e599b312.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 163, 638, 444, 669.6666666666666 ], "spans": [], "index": 22 }, { "bbox": [ 163, 669.6666666666666, 444, 701.3333333333333 ], "spans": [], "index": 23 }, { "bbox": [ 163, 701.3333333333333, 444, 732.9999999999999 ], "spans": [], "index": 24 } ] } ], "index": 21.75 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 761 ], "spans": [ { "bbox": [ 302, 750, 309, 761 ], "score": 1.0, "content": "8", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 309, 39 ], "spans": [ { "bbox": [ 106, 25, 309, 39 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 81, 478, 246 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 81, 478, 246 ], "group_id": 0, "lines": [ { "bbox": [ 132, 81, 478, 246 ], "spans": [ { "bbox": [ 132, 81, 478, 246 ], "score": 0.968, "type": "image", "image_path": "91696fc85e55662030167a35ef4aae436fa0857c2b3a62e62f0e14e5967749b8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 81, 478, 136.0 ], "spans": [], "index": 0 }, { "bbox": [ 132, 136.0, 478, 191.0 ], "spans": [], "index": 1 }, { "bbox": [ 132, 191.0, 478, 246.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 258, 506, 292 ], "group_id": 0, "lines": [ { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 349, 271 ], "score": 1.0, "content": "Figure 5: Ablation study. (a) The baseline StarGAN v2. (b)", "type": "text" }, { "bbox": [ 349, 260, 358, 269 ], "score": 0.66, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 359, 258, 471, 271 ], "score": 1.0, "content": "Style-Content branches. (c)", "type": "text" }, { "bbox": [ 472, 260, 481, 269 ], "score": 0.72, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 481, 258, 505, 271 ], "score": 1.0, "content": "Fore-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 269, 506, 282 ], "spans": [ { "bbox": [ 105, 269, 190, 282 ], "score": 1.0, "content": "ground classifier. (d)", "type": "text" }, { "bbox": [ 190, 270, 199, 279 ], "score": 0.48, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 199, 269, 302, 282 ], "score": 1.0, "content": "Background classifier. (e)", "type": "text" }, { "bbox": [ 302, 271, 310, 279 ], "score": 0.64, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 311, 269, 506, 282 ], "score": 1.0, "content": "Separately decoding foreground and background", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 280, 502, 293 ], "spans": [ { "bbox": [ 105, 280, 117, 293 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 117, 281, 126, 290 ], "score": 0.53, "content": "G", "type": "inline_equation" }, { "bbox": [ 126, 280, 141, 293 ], "score": 1.0, "content": ". (f)", "type": "text" }, { "bbox": [ 142, 281, 150, 290 ], "score": 0.69, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 151, 280, 341, 293 ], "score": 1.0, "content": "Anchor foreground domain (e.g. Normal). (g)", "type": "text" }, { "bbox": [ 341, 281, 349, 290 ], "score": 0.59, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 350, 280, 502, 293 ], "score": 1.0, "content": "Noise injection in Mapping Network.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 112, 297, 499, 427 ], "blocks": [ { "type": "image_body", "bbox": [ 112, 297, 499, 427 ], "group_id": 1, "lines": [ { "bbox": [ 112, 297, 499, 427 ], "spans": [ { "bbox": [ 112, 297, 499, 427 ], "score": 0.954, "type": "image", "image_path": "10b4554eb02cdbab48f0a3b29e5fed2318aee10f044aacf5703500d8bb21eef0.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 112, 297, 499, 340.3333333333333 ], "spans": [], "index": 6 }, { "bbox": [ 112, 340.3333333333333, 499, 383.66666666666663 ], "spans": [], "index": 7 }, { "bbox": [ 112, 383.66666666666663, 499, 426.99999999999994 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 105, 429, 505, 452 ], "group_id": 1, "lines": [ { "bbox": [ 106, 429, 505, 442 ], "spans": [ { "bbox": [ 106, 429, 505, 442 ], "score": 1.0, "content": "Figure 6: Visual effect of randomly sampled style codes on fixed pairs of reference background", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 439, 280, 454 ], "spans": [ { "bbox": [ 105, 439, 280, 454 ], "score": 1.0, "content": "(Source) and foreground (Content) images.", "type": "text" } ], "index": 10 } ], "index": 9.5 } ], "index": 8.25 }, { "type": "title", "bbox": [ 108, 477, 289, 488 ], "lines": [ { "bbox": [ 106, 476, 291, 490 ], "spans": [ { "bbox": [ 106, 476, 291, 490 ], "score": 1.0, "content": "4.2 DT-GAN FOR DATA AUGMENTATION", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 107, 500, 505, 533 ], "lines": [ { "bbox": [ 106, 500, 505, 512 ], "spans": [ { "bbox": [ 106, 500, 505, 512 ], "score": 1.0, "content": "We also evaluated our method as a data augmentation method for defect classification on the SDI", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 512, 504, 523 ], "spans": [ { "bbox": [ 106, 512, 504, 523 ], "score": 1.0, "content": "dataset. We defined one task ‘general’, where the classifier was trained on images from all products", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 522, 440, 534 ], "spans": [ { "bbox": [ 105, 522, 440, 534 ], "score": 1.0, "content": "at once, while task ‘single product’ only used the subset of images for one product.", "type": "text" } ], "index": 14 } ], "index": 13, "bbox_fs": [ 105, 500, 505, 534 ] }, { "type": "text", "bbox": [ 107, 539, 505, 595 ], "lines": [ { "bbox": [ 105, 537, 505, 552 ], "spans": [ { "bbox": [ 105, 537, 505, 552 ], "score": 1.0, "content": "Besides, we incrementally varied the amount of real Normal data available for classifier training:", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 550, 505, 563 ], "spans": [ { "bbox": [ 105, 550, 505, 563 ], "score": 1.0, "content": "4500, 6600, 12000 and 18600. In the case of defective images, all of them were always used due to", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 560, 505, 572 ], "spans": [ { "bbox": [ 106, 560, 505, 572 ], "score": 1.0, "content": "the small amount unless otherwise specified. As backbone we used a ResNet-50 (He et al., 2016a)", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 570, 505, 586 ], "spans": [ { "bbox": [ 105, 570, 505, 586 ], "score": 1.0, "content": "with ImageNet pretrained weights. For experiments with synthetic data, we attached an auxiliary", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 583, 493, 595 ], "spans": [ { "bbox": [ 106, 583, 493, 595 ], "score": 1.0, "content": "domain classifier to the network through a Gradient Reversal Layer (Ganin & Lempitsky, 2015).", "type": "text" } ], "index": 19 } ], "index": 17, "bbox_fs": [ 105, 537, 505, 595 ] }, { "type": "table", "bbox": [ 163, 638, 444, 733 ], "blocks": [ { "type": "table_caption", "bbox": [ 105, 615, 504, 638 ], "group_id": 0, "lines": [ { "bbox": [ 106, 615, 505, 628 ], "spans": [ { "bbox": [ 106, 615, 505, 628 ], "score": 1.0, "content": "Table 2: Quantitative comparison of the baseline methods on defect classification task at the scale", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 626, 469, 639 ], "spans": [ { "bbox": [ 105, 626, 392, 639 ], "score": 1.0, "content": "of 12000 images/class. The reported values are the achieved error rates", "type": "text" }, { "bbox": [ 392, 627, 408, 637 ], "score": 0.78, "content": "( \\% )", "type": "inline_equation" }, { "bbox": [ 408, 626, 469, 639 ], "score": 1.0, "content": "over five runs.", "type": "text" } ], "index": 21 } ], "index": 20.5 }, { "type": "table_body", "bbox": [ 163, 638, 444, 733 ], "group_id": 0, "lines": [ { "bbox": [ 163, 638, 444, 733 ], "spans": [ { "bbox": [ 163, 638, 444, 733 ], "score": 0.899, "html": "
MethodResNet-50EfficientNet-b4
No-Aug21.64±1.2412.06±0.64
Trad-Aug12.58±0.819.33±0.73
Mokady (2020)11.11±1.1913.26±1.13
StarGAN v213.07±1.3012.25±0.79
StyleGAN v211.55±1.7911.68±0.76
BigGAN+DiffAug11.45±0.6112.06±0.50
Ours9.9±0.699.14±1.02
", "type": "table", "image_path": "928c69ed8d652dba082f99aeef3e7de87cc055491f1db2d2d419fae8e599b312.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 163, 638, 444, 669.6666666666666 ], "spans": [], "index": 22 }, { "bbox": [ 163, 669.6666666666666, 444, 701.3333333333333 ], "spans": [], "index": 23 }, { "bbox": [ 163, 701.3333333333333, 444, 732.9999999999999 ], "spans": [], "index": 24 } ] } ], "index": 21.75 } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 137 ], "lines": [ { "bbox": [ 105, 81, 505, 95 ], "spans": [ { "bbox": [ 105, 81, 505, 95 ], "score": 1.0, "content": "Since the SDI dataset is highly imbalanced, we oversampled the minority classes (Ling et al., 1998)", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 94, 505, 106 ], "spans": [ { "bbox": [ 106, 94, 505, 106 ], "score": 1.0, "content": "unless the data was balanced through synthetic images. Additionally, we always applied traditional", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 105, 504, 117 ], "spans": [ { "bbox": [ 106, 105, 504, 117 ], "score": 1.0, "content": "data augmentation techniques like random horizontal flips, jittering and lighting (Shorten & Khosh-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 114, 506, 129 ], "spans": [ { "bbox": [ 105, 114, 506, 129 ], "score": 1.0, "content": "goftaar, 2019) except where noted. All following results were evaluated by the achieved error rates", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 126, 279, 138 ], "spans": [ { "bbox": [ 106, 126, 279, 138 ], "score": 1.0, "content": "over five runs with different random seeds.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "text", "bbox": [ 106, 143, 505, 210 ], "lines": [ { "bbox": [ 105, 142, 505, 156 ], "spans": [ { "bbox": [ 105, 142, 505, 156 ], "score": 1.0, "content": "Effectiveness of synthetic data. We first compare classifier performance for no augmentation (No-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 155, 505, 166 ], "spans": [ { "bbox": [ 106, 155, 505, 166 ], "score": 1.0, "content": "Aug), traditional data augmentation (Trad-Aug), and a combination of traditional augmentation", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 166, 505, 177 ], "spans": [ { "bbox": [ 106, 166, 505, 177 ], "score": 1.0, "content": "with synthetic images for GAN methods including DT-GAN. We also introduce a stronger back-", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 176, 505, 188 ], "spans": [ { "bbox": [ 106, 176, 505, 188 ], "score": 1.0, "content": "bone, EfficientNet-b4 (Tan & Le, 2019), to demonstrate that our results are not confined to a specific", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 187, 505, 199 ], "spans": [ { "bbox": [ 105, 187, 505, 199 ], "score": 1.0, "content": "network. Table 2 shows that our method is the only one that improves performance for both back-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 199, 504, 210 ], "spans": [ { "bbox": [ 106, 199, 504, 210 ], "score": 1.0, "content": "bones, presumably due to the combination of high visual image quality and diversity in our samples.", "type": "text" } ], "index": 10 } ], "index": 7.5 }, { "type": "table", "bbox": [ 173, 259, 436, 339 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 226, 504, 259 ], "group_id": 0, "lines": [ { "bbox": [ 105, 225, 506, 239 ], "spans": [ { "bbox": [ 105, 225, 506, 239 ], "score": 1.0, "content": "Table 3: Experimental results on using different amount of synthetic images generated by DT-GAN", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 237, 505, 249 ], "spans": [ { "bbox": [ 105, 237, 505, 249 ], "score": 1.0, "content": "to train classifiers. The left-most column stands for number of samples per class to be classified.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 247, 489, 262 ], "spans": [ { "bbox": [ 105, 247, 489, 262 ], "score": 1.0, "content": "The training set of the baselines is balanced by oversampling while ours is by synthetic images.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "table_body", "bbox": [ 173, 259, 436, 339 ], "group_id": 0, "lines": [ { "bbox": [ 173, 259, 436, 339 ], "spans": [ { "bbox": [ 173, 259, 436, 339 ], "score": 0.977, "html": "
Dataset Size20AAll
Trad-AugOursTrad-AugOurs
450015.55±0.6314.28±1.2512.75±0.6111.04±0.76
660016.69±0.7614.41±3.1213.07±1.5710.60±0.48
1200016.95±1.0214.22±1.5312.05±0.819.90±0.69
1860016.12±2.1915.36±0.8612.37±0.3210.21±0.96
", "type": "table", "image_path": "8d583e7701fa68183a5b7e4a5236d7a737bd6b648588956f2a2416af51c4e12b.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 173, 259, 436, 285.6666666666667 ], "spans": [], "index": 14 }, { "bbox": [ 173, 285.6666666666667, 436, 312.33333333333337 ], "spans": [], "index": 15 }, { "bbox": [ 173, 312.33333333333337, 436, 339.00000000000006 ], "spans": [], "index": 16 } ] } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 347, 505, 414 ], "lines": [ { "bbox": [ 105, 347, 505, 361 ], "spans": [ { "bbox": [ 105, 347, 505, 361 ], "score": 1.0, "content": "Impact of dataset size. Motivated by the limited availability of data in real-world production sce-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 358, 506, 371 ], "spans": [ { "bbox": [ 105, 358, 506, 371 ], "score": 1.0, "content": "narios, we therefore evaluated DT-GAN for data augmentation on a subset of the full SDI dataset", "type": "text" } ], "index": 18 }, { "bbox": [ 104, 368, 506, 383 ], "spans": [ { "bbox": [ 104, 368, 506, 383 ], "score": 1.0, "content": "(All), which only contains 20 defective samples in product A for each defect type (20A). In this case,", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 380, 505, 393 ], "spans": [ { "bbox": [ 105, 380, 505, 393 ], "score": 1.0, "content": "DT-GAN was also trained on the reduced subset. As shown in Table 3, there is a clear improvement", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 392, 506, 405 ], "spans": [ { "bbox": [ 105, 392, 506, 405 ], "score": 1.0, "content": "when synthetic images from DT-GAN are used as data augmentation, even for the extremely limited", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 402, 455, 415 ], "spans": [ { "bbox": [ 105, 402, 455, 415 ], "score": 1.0, "content": "data subset. Further results on single product classifiers can be found in Appendix E.1.", "type": "text" } ], "index": 22 } ], "index": 19.5 }, { "type": "table", "bbox": [ 155, 450, 454, 505 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 428, 504, 450 ], "group_id": 1, "lines": [ { "bbox": [ 105, 426, 506, 441 ], "spans": [ { "bbox": [ 105, 426, 506, 441 ], "score": 1.0, "content": "Table 4: Cross-domain effect on single product classifiers trained with reference-guided synthetic", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 438, 277, 451 ], "spans": [ { "bbox": [ 106, 438, 277, 451 ], "score": 1.0, "content": "images at the scale of 12000 images/class.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "table_body", "bbox": [ 155, 450, 454, 505 ], "group_id": 1, "lines": [ { "bbox": [ 155, 450, 454, 505 ], "spans": [ { "bbox": [ 155, 450, 454, 505 ], "score": 0.977, "html": "
Trad-AugvAVBvCvABC
A13.81±2.3611.81±2.6512.72±2.8711.99±1.6311.09±3.49
B6.80±1.646.40±1.346.60±1.526.59±1.345.60±1.34
C16.57±3.2013.14±2.8111.23±0.8014.85±1.7311.42±0.96
", "type": "table", "image_path": "974a838e3b729a382b7808545ffbfcb351041081afc0d6165f506ec8cd4fa186.jpg" } ] } ], "index": 26, "virtual_lines": [ { "bbox": [ 155, 450, 454, 468.3333333333333 ], "spans": [], "index": 25 }, { "bbox": [ 155, 468.3333333333333, 454, 486.66666666666663 ], "spans": [], "index": 26 }, { "bbox": [ 155, 486.66666666666663, 454, 504.99999999999994 ], "spans": [], "index": 27 } ] } ], "index": 24.75 }, { "type": "text", "bbox": [ 106, 513, 505, 591 ], "lines": [ { "bbox": [ 106, 512, 505, 527 ], "spans": [ { "bbox": [ 106, 512, 505, 527 ], "score": 1.0, "content": "Cross-domain effect. We hypothesized that limited data can be counteracted by transferring de-", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 525, 504, 536 ], "spans": [ { "bbox": [ 106, 525, 504, 536 ], "score": 1.0, "content": "fects across multiple background products, if there are at least some defects that occur on multiple", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 535, 505, 549 ], "spans": [ { "bbox": [ 105, 535, 505, 549 ], "score": 1.0, "content": "products (See Appendix E.1 for further discussion). We tested this approach by comparing the per-", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 547, 504, 559 ], "spans": [ { "bbox": [ 106, 547, 504, 559 ], "score": 1.0, "content": "formance of classifiers trained on synthetic images with defects from a specific source (vA, vB, vC)", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 558, 505, 569 ], "spans": [ { "bbox": [ 106, 558, 505, 569 ], "score": 1.0, "content": "to classifiers trained on images with defects from all products (vABC). As we can see in Table 4,", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 569, 505, 581 ], "spans": [ { "bbox": [ 106, 569, 505, 581 ], "score": 1.0, "content": "the best performances are reached by the models that take over defects from other products. We", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 579, 390, 592 ], "spans": [ { "bbox": [ 106, 579, 390, 592 ], "score": 1.0, "content": "interpret this as support for our hypothesis and its practical usefulness.", "type": "text" } ], "index": 34 } ], "index": 31 }, { "type": "title", "bbox": [ 108, 607, 195, 620 ], "lines": [ { "bbox": [ 104, 604, 198, 624 ], "spans": [ { "bbox": [ 104, 604, 198, 624 ], "score": 1.0, "content": "5 CONCLUSION", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 106, 632, 505, 732 ], "lines": [ { "bbox": [ 105, 631, 506, 647 ], "spans": [ { "bbox": [ 105, 631, 506, 647 ], "score": 1.0, "content": "We propose a novel method, DT-GAN, which allows diverse defect synthesis both by generating", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 643, 506, 657 ], "spans": [ { "bbox": [ 105, 643, 506, 657 ], "score": 1.0, "content": "from randomly sampled noise and by following the guidance of given reference images. Due to", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 654, 506, 669 ], "spans": [ { "bbox": [ 105, 654, 506, 669 ], "score": 1.0, "content": "explicit style-content separation and FG/BG disentanglement, DT-GAN achieves higher image", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 664, 506, 680 ], "spans": [ { "bbox": [ 105, 664, 506, 680 ], "score": 1.0, "content": "fidelity, better variance in defects and full control over background and foreground while being", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 677, 506, 689 ], "spans": [ { "bbox": [ 106, 677, 506, 689 ], "score": 1.0, "content": "sample-efficient. We demonstrated the feasibility and benefits of DT-GAN on a real industrial defect", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 687, 506, 700 ], "spans": [ { "bbox": [ 106, 687, 506, 700 ], "score": 1.0, "content": "classification task and the results show our method provides consistent gains even with limited data", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "and boosts the performance of classifiers compared to state-of-the-art image synthesis methods. For", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "future investigation, we aim to represent defects more explicitly (e.g., localization) to improve the", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 720, 467, 734 ], "spans": [ { "bbox": [ 106, 720, 467, 734 ], "score": 1.0, "content": "explainability of the model and also enhance the model transferability to unseen products.", "type": "text" } ], "index": 44 } ], "index": 40 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 137 ], "lines": [ { "bbox": [ 105, 81, 505, 95 ], "spans": [ { "bbox": [ 105, 81, 505, 95 ], "score": 1.0, "content": "Since the SDI dataset is highly imbalanced, we oversampled the minority classes (Ling et al., 1998)", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 94, 505, 106 ], "spans": [ { "bbox": [ 106, 94, 505, 106 ], "score": 1.0, "content": "unless the data was balanced through synthetic images. Additionally, we always applied traditional", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 105, 504, 117 ], "spans": [ { "bbox": [ 106, 105, 504, 117 ], "score": 1.0, "content": "data augmentation techniques like random horizontal flips, jittering and lighting (Shorten & Khosh-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 114, 506, 129 ], "spans": [ { "bbox": [ 105, 114, 506, 129 ], "score": 1.0, "content": "goftaar, 2019) except where noted. All following results were evaluated by the achieved error rates", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 126, 279, 138 ], "spans": [ { "bbox": [ 106, 126, 279, 138 ], "score": 1.0, "content": "over five runs with different random seeds.", "type": "text" } ], "index": 4 } ], "index": 2, "bbox_fs": [ 105, 81, 506, 138 ] }, { "type": "text", "bbox": [ 106, 143, 505, 210 ], "lines": [ { "bbox": [ 105, 142, 505, 156 ], "spans": [ { "bbox": [ 105, 142, 505, 156 ], "score": 1.0, "content": "Effectiveness of synthetic data. We first compare classifier performance for no augmentation (No-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 155, 505, 166 ], "spans": [ { "bbox": [ 106, 155, 505, 166 ], "score": 1.0, "content": "Aug), traditional data augmentation (Trad-Aug), and a combination of traditional augmentation", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 166, 505, 177 ], "spans": [ { "bbox": [ 106, 166, 505, 177 ], "score": 1.0, "content": "with synthetic images for GAN methods including DT-GAN. We also introduce a stronger back-", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 176, 505, 188 ], "spans": [ { "bbox": [ 106, 176, 505, 188 ], "score": 1.0, "content": "bone, EfficientNet-b4 (Tan & Le, 2019), to demonstrate that our results are not confined to a specific", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 187, 505, 199 ], "spans": [ { "bbox": [ 105, 187, 505, 199 ], "score": 1.0, "content": "network. Table 2 shows that our method is the only one that improves performance for both back-", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 199, 504, 210 ], "spans": [ { "bbox": [ 106, 199, 504, 210 ], "score": 1.0, "content": "bones, presumably due to the combination of high visual image quality and diversity in our samples.", "type": "text" } ], "index": 10 } ], "index": 7.5, "bbox_fs": [ 105, 142, 505, 210 ] }, { "type": "table", "bbox": [ 173, 259, 436, 339 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 226, 504, 259 ], "group_id": 0, "lines": [ { "bbox": [ 105, 225, 506, 239 ], "spans": [ { "bbox": [ 105, 225, 506, 239 ], "score": 1.0, "content": "Table 3: Experimental results on using different amount of synthetic images generated by DT-GAN", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 237, 505, 249 ], "spans": [ { "bbox": [ 105, 237, 505, 249 ], "score": 1.0, "content": "to train classifiers. The left-most column stands for number of samples per class to be classified.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 247, 489, 262 ], "spans": [ { "bbox": [ 105, 247, 489, 262 ], "score": 1.0, "content": "The training set of the baselines is balanced by oversampling while ours is by synthetic images.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "table_body", "bbox": [ 173, 259, 436, 339 ], "group_id": 0, "lines": [ { "bbox": [ 173, 259, 436, 339 ], "spans": [ { "bbox": [ 173, 259, 436, 339 ], "score": 0.977, "html": "
Dataset Size20AAll
Trad-AugOursTrad-AugOurs
450015.55±0.6314.28±1.2512.75±0.6111.04±0.76
660016.69±0.7614.41±3.1213.07±1.5710.60±0.48
1200016.95±1.0214.22±1.5312.05±0.819.90±0.69
1860016.12±2.1915.36±0.8612.37±0.3210.21±0.96
", "type": "table", "image_path": "8d583e7701fa68183a5b7e4a5236d7a737bd6b648588956f2a2416af51c4e12b.jpg" } ] } ], "index": 15, "virtual_lines": [ { "bbox": [ 173, 259, 436, 285.6666666666667 ], "spans": [], "index": 14 }, { "bbox": [ 173, 285.6666666666667, 436, 312.33333333333337 ], "spans": [], "index": 15 }, { "bbox": [ 173, 312.33333333333337, 436, 339.00000000000006 ], "spans": [], "index": 16 } ] } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 347, 505, 414 ], "lines": [ { "bbox": [ 105, 347, 505, 361 ], "spans": [ { "bbox": [ 105, 347, 505, 361 ], "score": 1.0, "content": "Impact of dataset size. Motivated by the limited availability of data in real-world production sce-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 358, 506, 371 ], "spans": [ { "bbox": [ 105, 358, 506, 371 ], "score": 1.0, "content": "narios, we therefore evaluated DT-GAN for data augmentation on a subset of the full SDI dataset", "type": "text" } ], "index": 18 }, { "bbox": [ 104, 368, 506, 383 ], "spans": [ { "bbox": [ 104, 368, 506, 383 ], "score": 1.0, "content": "(All), which only contains 20 defective samples in product A for each defect type (20A). In this case,", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 380, 505, 393 ], "spans": [ { "bbox": [ 105, 380, 505, 393 ], "score": 1.0, "content": "DT-GAN was also trained on the reduced subset. As shown in Table 3, there is a clear improvement", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 392, 506, 405 ], "spans": [ { "bbox": [ 105, 392, 506, 405 ], "score": 1.0, "content": "when synthetic images from DT-GAN are used as data augmentation, even for the extremely limited", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 402, 455, 415 ], "spans": [ { "bbox": [ 105, 402, 455, 415 ], "score": 1.0, "content": "data subset. Further results on single product classifiers can be found in Appendix E.1.", "type": "text" } ], "index": 22 } ], "index": 19.5, "bbox_fs": [ 104, 347, 506, 415 ] }, { "type": "table", "bbox": [ 155, 450, 454, 505 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 428, 504, 450 ], "group_id": 1, "lines": [ { "bbox": [ 105, 426, 506, 441 ], "spans": [ { "bbox": [ 105, 426, 506, 441 ], "score": 1.0, "content": "Table 4: Cross-domain effect on single product classifiers trained with reference-guided synthetic", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 438, 277, 451 ], "spans": [ { "bbox": [ 106, 438, 277, 451 ], "score": 1.0, "content": "images at the scale of 12000 images/class.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "table_body", "bbox": [ 155, 450, 454, 505 ], "group_id": 1, "lines": [ { "bbox": [ 155, 450, 454, 505 ], "spans": [ { "bbox": [ 155, 450, 454, 505 ], "score": 0.977, "html": "
Trad-AugvAVBvCvABC
A13.81±2.3611.81±2.6512.72±2.8711.99±1.6311.09±3.49
B6.80±1.646.40±1.346.60±1.526.59±1.345.60±1.34
C16.57±3.2013.14±2.8111.23±0.8014.85±1.7311.42±0.96
", "type": "table", "image_path": "974a838e3b729a382b7808545ffbfcb351041081afc0d6165f506ec8cd4fa186.jpg" } ] } ], "index": 26, "virtual_lines": [ { "bbox": [ 155, 450, 454, 468.3333333333333 ], "spans": [], "index": 25 }, { "bbox": [ 155, 468.3333333333333, 454, 486.66666666666663 ], "spans": [], "index": 26 }, { "bbox": [ 155, 486.66666666666663, 454, 504.99999999999994 ], "spans": [], "index": 27 } ] } ], "index": 24.75 }, { "type": "text", "bbox": [ 106, 513, 505, 591 ], "lines": [ { "bbox": [ 106, 512, 505, 527 ], "spans": [ { "bbox": [ 106, 512, 505, 527 ], "score": 1.0, "content": "Cross-domain effect. We hypothesized that limited data can be counteracted by transferring de-", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 525, 504, 536 ], "spans": [ { "bbox": [ 106, 525, 504, 536 ], "score": 1.0, "content": "fects across multiple background products, if there are at least some defects that occur on multiple", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 535, 505, 549 ], "spans": [ { "bbox": [ 105, 535, 505, 549 ], "score": 1.0, "content": "products (See Appendix E.1 for further discussion). We tested this approach by comparing the per-", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 547, 504, 559 ], "spans": [ { "bbox": [ 106, 547, 504, 559 ], "score": 1.0, "content": "formance of classifiers trained on synthetic images with defects from a specific source (vA, vB, vC)", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 558, 505, 569 ], "spans": [ { "bbox": [ 106, 558, 505, 569 ], "score": 1.0, "content": "to classifiers trained on images with defects from all products (vABC). As we can see in Table 4,", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 569, 505, 581 ], "spans": [ { "bbox": [ 106, 569, 505, 581 ], "score": 1.0, "content": "the best performances are reached by the models that take over defects from other products. We", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 579, 390, 592 ], "spans": [ { "bbox": [ 106, 579, 390, 592 ], "score": 1.0, "content": "interpret this as support for our hypothesis and its practical usefulness.", "type": "text" } ], "index": 34 } ], "index": 31, "bbox_fs": [ 105, 512, 505, 592 ] }, { "type": "title", "bbox": [ 108, 607, 195, 620 ], "lines": [ { "bbox": [ 104, 604, 198, 624 ], "spans": [ { "bbox": [ 104, 604, 198, 624 ], "score": 1.0, "content": "5 CONCLUSION", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 106, 632, 505, 732 ], "lines": [ { "bbox": [ 105, 631, 506, 647 ], "spans": [ { "bbox": [ 105, 631, 506, 647 ], "score": 1.0, "content": "We propose a novel method, DT-GAN, which allows diverse defect synthesis both by generating", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 643, 506, 657 ], "spans": [ { "bbox": [ 105, 643, 506, 657 ], "score": 1.0, "content": "from randomly sampled noise and by following the guidance of given reference images. Due to", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 654, 506, 669 ], "spans": [ { "bbox": [ 105, 654, 506, 669 ], "score": 1.0, "content": "explicit style-content separation and FG/BG disentanglement, DT-GAN achieves higher image", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 664, 506, 680 ], "spans": [ { "bbox": [ 105, 664, 506, 680 ], "score": 1.0, "content": "fidelity, better variance in defects and full control over background and foreground while being", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 677, 506, 689 ], "spans": [ { "bbox": [ 106, 677, 506, 689 ], "score": 1.0, "content": "sample-efficient. We demonstrated the feasibility and benefits of DT-GAN on a real industrial defect", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 687, 506, 700 ], "spans": [ { "bbox": [ 106, 687, 506, 700 ], "score": 1.0, "content": "classification task and the results show our method provides consistent gains even with limited data", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "and boosts the performance of classifiers compared to state-of-the-art image synthesis methods. For", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "future investigation, we aim to represent defects more explicitly (e.g., localization) to improve the", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 720, 467, 734 ], "spans": [ { "bbox": [ 106, 720, 467, 734 ], "score": 1.0, "content": "explainability of the model and also enhance the model transferability to unseen products.", "type": "text" } ], "index": 44 } ], "index": 40, "bbox_fs": [ 105, 631, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 82, 242, 94 ], "lines": [ { "bbox": [ 106, 83, 243, 94 ], "spans": [ { "bbox": [ 106, 83, 243, 94 ], "score": 1.0, "content": "REPRODUCIBILITY STATEMENT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 101, 504, 135 ], "lines": [ { "bbox": [ 106, 101, 505, 113 ], "spans": [ { "bbox": [ 106, 101, 505, 113 ], "score": 1.0, "content": "We aim for full reproducibility by publishing the source code and dataset with the final version of", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 113, 505, 125 ], "spans": [ { "bbox": [ 106, 113, 505, 125 ], "score": 1.0, "content": "the paper. Besides, we provide descriptions of the training details in Appendix B, the evaluation", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 123, 371, 136 ], "spans": [ { "bbox": [ 105, 123, 371, 136 ], "score": 1.0, "content": "setup in Appendix C and the network architecture in Appendix D.", "type": "text" } ], "index": 3 } ], "index": 2 }, { "type": "text", "bbox": [ 102, 134, 506, 736 ], "lines": [ { "bbox": [ 106, 151, 176, 165 ], "spans": [ { "bbox": [ 106, 151, 176, 165 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 170, 505, 183 ], "spans": [ { "bbox": [ 105, 170, 505, 183 ], "score": 1.0, "content": "Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a com-", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 180, 506, 194 ], "spans": [ { "bbox": [ 114, 180, 506, 194 ], "score": 1.0, "content": "prehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Con-", "type": "text" } ], "index": 6 }, { "bbox": [ 114, 191, 506, 205 ], "spans": [ { "bbox": [ 114, 191, 506, 205 ], "score": 1.0, "content": "ference on Computer Vision and Pattern Recognition (CVPR), pp. 9584–9592, 2019. doi:", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 201, 232, 215 ], "spans": [ { "bbox": [ 115, 201, 232, 215 ], "score": 1.0, "content": "10.1109/CVPR.2019.00982.", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 220, 505, 234 ], "spans": [ { "bbox": [ 105, 220, 505, 234 ], "score": 1.0, "content": "Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and A. Gretton. Demystifying MMD", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 232, 270, 244 ], "spans": [ { "bbox": [ 115, 232, 270, 244 ], "score": 1.0, "content": "GANs. ArXiv, abs/1801.01401, 2018.", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 248, 505, 264 ], "spans": [ { "bbox": [ 104, 248, 505, 264 ], "score": 1.0, "content": "Andrew Brock, Jeff Donahue, and K. Simonyan. Large scale gan training for high fidelity natural", "type": "text" } ], "index": 11 }, { "bbox": [ 114, 261, 308, 274 ], "spans": [ { "bbox": [ 114, 261, 308, 274 ], "score": 1.0, "content": "image synthesis. ArXiv, abs/1809.11096, 2019.", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 278, 505, 293 ], "spans": [ { "bbox": [ 106, 278, 505, 293 ], "score": 1.0, "content": "Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 290, 505, 303 ], "spans": [ { "bbox": [ 116, 290, 505, 303 ], "score": 1.0, "content": "cnn for object detection in the wild. 2018 IEEE/CVF Conference on Computer Vision and Pattern", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 300, 261, 314 ], "spans": [ { "bbox": [ 115, 300, 261, 314 ], "score": 1.0, "content": "Recognition, pp. 3339–3348, 2018.", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 318, 506, 333 ], "spans": [ { "bbox": [ 105, 318, 506, 333 ], "score": 1.0, "content": "Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis", "type": "text" } ], "index": 16 }, { "bbox": [ 114, 329, 506, 344 ], "spans": [ { "bbox": [ 114, 329, 506, 344 ], "score": 1.0, "content": "for multiple domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 342, 281, 354 ], "spans": [ { "bbox": [ 116, 342, 281, 354 ], "score": 1.0, "content": "Pattern Recognition (CVPR), June 2020.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 360, 505, 372 ], "spans": [ { "bbox": [ 105, 360, 505, 372 ], "score": 1.0, "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi-", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 369, 506, 385 ], "spans": [ { "bbox": [ 115, 369, 506, 385 ], "score": 1.0, "content": "erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,", "type": "text" } ], "index": 20 }, { "bbox": [ 114, 381, 218, 394 ], "spans": [ { "bbox": [ 114, 381, 218, 394 ], "score": 1.0, "content": "pp. 248–255. Ieee, 2009.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 398, 505, 413 ], "spans": [ { "bbox": [ 105, 398, 505, 413 ], "score": 1.0, "content": "Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In", "type": "text" } ], "index": 22 }, { "bbox": [ 114, 409, 505, 424 ], "spans": [ { "bbox": [ 114, 409, 505, 424 ], "score": 1.0, "content": "Francis Bach and David Blei (eds.), Proceedings of the 32nd International Conference on Ma-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 420, 505, 436 ], "spans": [ { "bbox": [ 115, 420, 505, 436 ], "score": 1.0, "content": "chine Learning, volume 37 of Proceedings of Machine Learning Research, pp. 1180–1189,", "type": "text" } ], "index": 24 }, { "bbox": [ 114, 431, 505, 446 ], "spans": [ { "bbox": [ 114, 431, 505, 446 ], "score": 1.0, "content": "Lille, France, 07–09 Jul 2015. PMLR. URL https://proceedings.mlr.press/v37/", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 444, 194, 456 ], "spans": [ { "bbox": [ 115, 444, 194, 456 ], "score": 1.0, "content": "ganin15.html.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 462, 505, 474 ], "spans": [ { "bbox": [ 105, 462, 505, 474 ], "score": 1.0, "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", "type": "text" } ], "index": 27 }, { "bbox": [ 114, 472, 479, 486 ], "spans": [ { "bbox": [ 114, 472, 479, 486 ], "score": 1.0, "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In NeurIPS, 2014.", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 491, 504, 504 ], "spans": [ { "bbox": [ 106, 491, 504, 504 ], "score": 1.0, "content": "Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recogni-", "type": "text" } ], "index": 29 }, { "bbox": [ 114, 500, 506, 515 ], "spans": [ { "bbox": [ 114, 500, 506, 515 ], "score": 1.0, "content": "tion. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778,", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 511, 147, 527 ], "spans": [ { "bbox": [ 115, 511, 147, 527 ], "score": 1.0, "content": "2016a.", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 530, 506, 545 ], "spans": [ { "bbox": [ 105, 530, 394, 545 ], "score": 1.0, "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.", "type": "text" }, { "bbox": [ 406, 530, 506, 545 ], "score": 1.0, "content": "Identity mappings in", "type": "text" } ], "index": 32 }, { "bbox": [ 114, 541, 506, 556 ], "spans": [ { "bbox": [ 114, 541, 506, 556 ], "score": 1.0, "content": "deep residual networks, 2016b. URL http://arxiv.org/abs/1603.05027. cite", "type": "text" } ], "index": 33 }, { "bbox": [ 114, 551, 343, 567 ], "spans": [ { "bbox": [ 114, 551, 343, 567 ], "score": 1.0, "content": "arxiv:1603.05027Comment: ECCV 2016 camera-ready.", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 571, 505, 584 ], "spans": [ { "bbox": [ 105, 571, 505, 584 ], "score": 1.0, "content": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and S. Hochreiter. Gans", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 581, 487, 596 ], "spans": [ { "bbox": [ 115, 581, 487, 596 ], "score": 1.0, "content": "trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017.", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 601, 505, 613 ], "spans": [ { "bbox": [ 105, 601, 505, 613 ], "score": 1.0, "content": "Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normal-", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 611, 506, 624 ], "spans": [ { "bbox": [ 115, 611, 506, 624 ], "score": 1.0, "content": "ization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 622, 144, 635 ], "spans": [ { "bbox": [ 115, 622, 144, 635 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 639, 505, 654 ], "spans": [ { "bbox": [ 104, 639, 505, 654 ], "score": 1.0, "content": "Xun Huang, Ming-Yu Liu, Serge J. Belongie, and J. Kautz. Multimodal unsupervised image-to-", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 652, 313, 664 ], "spans": [ { "bbox": [ 115, 652, 313, 664 ], "score": 1.0, "content": "image translation. ArXiv, abs/1804.04732, 2018.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 670, 505, 682 ], "spans": [ { "bbox": [ 105, 670, 505, 682 ], "score": 1.0, "content": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with", "type": "text" } ], "index": 42 }, { "bbox": [ 113, 678, 505, 696 ], "spans": [ { "bbox": [ 113, 678, 505, 696 ], "score": 1.0, "content": "conditional adversarial networks. 2017 IEEE Conference on Computer Vision and Pattern Recog-", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 691, 268, 704 ], "spans": [ { "bbox": [ 116, 691, 268, 704 ], "score": 1.0, "content": "nition (CVPR), pp. 5967–5976, 2017.", "type": "text" } ], "index": 44 }, { "bbox": [ 104, 708, 506, 723 ], "spans": [ { "bbox": [ 104, 708, 506, 723 ], "score": 1.0, "content": "Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and", "type": "text" } ], "index": 45 }, { "bbox": [ 114, 721, 211, 732 ], "spans": [ { "bbox": [ 114, 721, 211, 732 ], "score": 1.0, "content": "super-resolution, 2016.", "type": "text" } ], "index": 46 } ], "index": 25 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "10", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 82, 242, 94 ], "lines": [ { "bbox": [ 106, 83, 243, 94 ], "spans": [ { "bbox": [ 106, 83, 243, 94 ], "score": 1.0, "content": "REPRODUCIBILITY STATEMENT", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 101, 504, 135 ], "lines": [ { "bbox": [ 106, 101, 505, 113 ], "spans": [ { "bbox": [ 106, 101, 505, 113 ], "score": 1.0, "content": "We aim for full reproducibility by publishing the source code and dataset with the final version of", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 113, 505, 125 ], "spans": [ { "bbox": [ 106, 113, 505, 125 ], "score": 1.0, "content": "the paper. Besides, we provide descriptions of the training details in Appendix B, the evaluation", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 123, 371, 136 ], "spans": [ { "bbox": [ 105, 123, 371, 136 ], "score": 1.0, "content": "setup in Appendix C and the network architecture in Appendix D.", "type": "text" } ], "index": 3 } ], "index": 2, "bbox_fs": [ 105, 101, 505, 136 ] }, { "type": "list", "bbox": [ 102, 134, 506, 736 ], "lines": [ { "bbox": [ 106, 151, 176, 165 ], "spans": [ { "bbox": [ 106, 151, 176, 165 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 4, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 170, 505, 183 ], "spans": [ { "bbox": [ 105, 170, 505, 183 ], "score": 1.0, "content": "Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a com-", "type": "text" } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 114, 180, 506, 194 ], "spans": [ { "bbox": [ 114, 180, 506, 194 ], "score": 1.0, "content": "prehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Con-", "type": "text" } ], "index": 6 }, { "bbox": [ 114, 191, 506, 205 ], "spans": [ { "bbox": [ 114, 191, 506, 205 ], "score": 1.0, "content": "ference on Computer Vision and Pattern Recognition (CVPR), pp. 9584–9592, 2019. doi:", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 201, 232, 215 ], "spans": [ { "bbox": [ 115, 201, 232, 215 ], "score": 1.0, "content": "10.1109/CVPR.2019.00982.", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 105, 220, 505, 234 ], "spans": [ { "bbox": [ 105, 220, 505, 234 ], "score": 1.0, "content": "Mikolaj Binkowski, Danica J. Sutherland, Michael Arbel, and A. Gretton. Demystifying MMD", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 115, 232, 270, 244 ], "spans": [ { "bbox": [ 115, 232, 270, 244 ], "score": 1.0, "content": "GANs. ArXiv, abs/1801.01401, 2018.", "type": "text" } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 104, 248, 505, 264 ], "spans": [ { "bbox": [ 104, 248, 505, 264 ], "score": 1.0, "content": "Andrew Brock, Jeff Donahue, and K. Simonyan. Large scale gan training for high fidelity natural", "type": "text" } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 114, 261, 308, 274 ], "spans": [ { "bbox": [ 114, 261, 308, 274 ], "score": 1.0, "content": "image synthesis. ArXiv, abs/1809.11096, 2019.", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 106, 278, 505, 293 ], "spans": [ { "bbox": [ 106, 278, 505, 293 ], "score": 1.0, "content": "Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-", "type": "text" } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 116, 290, 505, 303 ], "spans": [ { "bbox": [ 116, 290, 505, 303 ], "score": 1.0, "content": "cnn for object detection in the wild. 2018 IEEE/CVF Conference on Computer Vision and Pattern", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 300, 261, 314 ], "spans": [ { "bbox": [ 115, 300, 261, 314 ], "score": 1.0, "content": "Recognition, pp. 3339–3348, 2018.", "type": "text" } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 105, 318, 506, 333 ], "spans": [ { "bbox": [ 105, 318, 506, 333 ], "score": 1.0, "content": "Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis", "type": "text" } ], "index": 16, "is_list_start_line": true }, { "bbox": [ 114, 329, 506, 344 ], "spans": [ { "bbox": [ 114, 329, 506, 344 ], "score": 1.0, "content": "for multiple domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 342, 281, 354 ], "spans": [ { "bbox": [ 116, 342, 281, 354 ], "score": 1.0, "content": "Pattern Recognition (CVPR), June 2020.", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 360, 505, 372 ], "spans": [ { "bbox": [ 105, 360, 505, 372 ], "score": 1.0, "content": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi-", "type": "text" } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 115, 369, 506, 385 ], "spans": [ { "bbox": [ 115, 369, 506, 385 ], "score": 1.0, "content": "erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition,", "type": "text" } ], "index": 20 }, { "bbox": [ 114, 381, 218, 394 ], "spans": [ { "bbox": [ 114, 381, 218, 394 ], "score": 1.0, "content": "pp. 248–255. Ieee, 2009.", "type": "text" } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 105, 398, 505, 413 ], "spans": [ { "bbox": [ 105, 398, 505, 413 ], "score": 1.0, "content": "Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In", "type": "text" } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 114, 409, 505, 424 ], "spans": [ { "bbox": [ 114, 409, 505, 424 ], "score": 1.0, "content": "Francis Bach and David Blei (eds.), Proceedings of the 32nd International Conference on Ma-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 420, 505, 436 ], "spans": [ { "bbox": [ 115, 420, 505, 436 ], "score": 1.0, "content": "chine Learning, volume 37 of Proceedings of Machine Learning Research, pp. 1180–1189,", "type": "text" } ], "index": 24 }, { "bbox": [ 114, 431, 505, 446 ], "spans": [ { "bbox": [ 114, 431, 505, 446 ], "score": 1.0, "content": "Lille, France, 07–09 Jul 2015. PMLR. URL https://proceedings.mlr.press/v37/", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 444, 194, 456 ], "spans": [ { "bbox": [ 115, 444, 194, 456 ], "score": 1.0, "content": "ganin15.html.", "type": "text" } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 105, 462, 505, 474 ], "spans": [ { "bbox": [ 105, 462, 505, 474 ], "score": 1.0, "content": "Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,", "type": "text" } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 114, 472, 479, 486 ], "spans": [ { "bbox": [ 114, 472, 479, 486 ], "score": 1.0, "content": "Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In NeurIPS, 2014.", "type": "text" } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 106, 491, 504, 504 ], "spans": [ { "bbox": [ 106, 491, 504, 504 ], "score": 1.0, "content": "Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recogni-", "type": "text" } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 114, 500, 506, 515 ], "spans": [ { "bbox": [ 114, 500, 506, 515 ], "score": 1.0, "content": "tion. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778,", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 511, 147, 527 ], "spans": [ { "bbox": [ 115, 511, 147, 527 ], "score": 1.0, "content": "2016a.", "type": "text" } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 105, 530, 506, 545 ], "spans": [ { "bbox": [ 105, 530, 394, 545 ], "score": 1.0, "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.", "type": "text" }, { "bbox": [ 406, 530, 506, 545 ], "score": 1.0, "content": "Identity mappings in", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 114, 541, 506, 556 ], "spans": [ { "bbox": [ 114, 541, 506, 556 ], "score": 1.0, "content": "deep residual networks, 2016b. URL http://arxiv.org/abs/1603.05027. cite", "type": "text" } ], "index": 33 }, { "bbox": [ 114, 551, 343, 567 ], "spans": [ { "bbox": [ 114, 551, 343, 567 ], "score": 1.0, "content": "arxiv:1603.05027Comment: ECCV 2016 camera-ready.", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 105, 571, 505, 584 ], "spans": [ { "bbox": [ 105, 571, 505, 584 ], "score": 1.0, "content": "Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and S. Hochreiter. Gans", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 115, 581, 487, 596 ], "spans": [ { "bbox": [ 115, 581, 487, 596 ], "score": 1.0, "content": "trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017.", "type": "text" } ], "index": 36, "is_list_end_line": true }, { "bbox": [ 105, 601, 505, 613 ], "spans": [ { "bbox": [ 105, 601, 505, 613 ], "score": 1.0, "content": "Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normal-", "type": "text" } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 115, 611, 506, 624 ], "spans": [ { "bbox": [ 115, 611, 506, 624 ], "score": 1.0, "content": "ization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 622, 144, 635 ], "spans": [ { "bbox": [ 115, 622, 144, 635 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 104, 639, 505, 654 ], "spans": [ { "bbox": [ 104, 639, 505, 654 ], "score": 1.0, "content": "Xun Huang, Ming-Yu Liu, Serge J. Belongie, and J. Kautz. Multimodal unsupervised image-to-", "type": "text" } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 115, 652, 313, 664 ], "spans": [ { "bbox": [ 115, 652, 313, 664 ], "score": 1.0, "content": "image translation. ArXiv, abs/1804.04732, 2018.", "type": "text" } ], "index": 41, "is_list_end_line": true }, { "bbox": [ 105, 670, 505, 682 ], "spans": [ { "bbox": [ 105, 670, 505, 682 ], "score": 1.0, "content": "Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with", "type": "text" } ], "index": 42, "is_list_start_line": true }, { "bbox": [ 113, 678, 505, 696 ], "spans": [ { "bbox": [ 113, 678, 505, 696 ], "score": 1.0, "content": "conditional adversarial networks. 2017 IEEE Conference on Computer Vision and Pattern Recog-", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 691, 268, 704 ], "spans": [ { "bbox": [ 116, 691, 268, 704 ], "score": 1.0, "content": "nition (CVPR), pp. 5967–5976, 2017.", "type": "text" } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 104, 708, 506, 723 ], "spans": [ { "bbox": [ 104, 708, 506, 723 ], "score": 1.0, "content": "Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and", "type": "text" } ], "index": 45, "is_list_start_line": true }, { "bbox": [ 114, 721, 211, 732 ], "spans": [ { "bbox": [ 114, 721, 211, 732 ], "score": 1.0, "content": "super-resolution, 2016.", "type": "text" } ], "index": 46, "is_list_end_line": true } ], "index": 25, "bbox_fs": [ 104, 151, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 105, 79, 506, 96 ], "spans": [ { "bbox": [ 105, 79, 506, 96 ], "score": 1.0, "content": "Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 347, 105 ], "spans": [ { "bbox": [ 115, 93, 347, 105 ], "score": 1.0, "content": "generative adversarial networks with limited data, 2020a.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 108, 113, 503, 147 ], "lines": [ { "bbox": [ 105, 112, 505, 127 ], "spans": [ { "bbox": [ 105, 112, 505, 127 ], "score": 1.0, "content": "Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyz-", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 125, 505, 137 ], "spans": [ { "bbox": [ 117, 125, 505, 137 ], "score": 1.0, "content": "ing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 135, 373, 149 ], "spans": [ { "bbox": [ 117, 135, 373, 149 ], "score": 1.0, "content": "Computer Vision and Pattern Recognition (CVPR), June 2020b.", "type": "text" } ], "index": 4 } ], "index": 3 }, { "type": "text", "bbox": [ 106, 155, 503, 178 ], "lines": [ { "bbox": [ 105, 154, 505, 169 ], "spans": [ { "bbox": [ 105, 154, 505, 169 ], "score": 1.0, "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech-", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 167, 366, 178 ], "spans": [ { "bbox": [ 115, 167, 366, 178 ], "score": 1.0, "content": "nical Report 0, University of Toronto, Toronto, Ontario, 2009.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "text", "bbox": [ 106, 186, 506, 220 ], "lines": [ { "bbox": [ 106, 186, 505, 199 ], "spans": [ { "bbox": [ 106, 186, 505, 199 ], "score": 1.0, "content": "Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro ´", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 198, 505, 210 ], "spans": [ { "bbox": [ 116, 198, 505, 210 ], "score": 1.0, "content": "Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 209, 436, 221 ], "spans": [ { "bbox": [ 116, 209, 436, 221 ], "score": 1.0, "content": "image super-resolution using a generative adversarial network. In CVPR, 2017.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "text", "bbox": [ 106, 227, 506, 262 ], "lines": [ { "bbox": [ 106, 228, 505, 240 ], "spans": [ { "bbox": [ 106, 228, 505, 240 ], "score": 1.0, "content": "Charles Ling, , Charles X. Ling, and Chenghui Li. Data mining for direct marketing: Problems and", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 240, 506, 252 ], "spans": [ { "bbox": [ 116, 240, 506, 252 ], "score": 1.0, "content": "solutions. In In Proceedings of the Fourth International Conference on Knowledge Discovery and", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 250, 336, 263 ], "spans": [ { "bbox": [ 116, 250, 336, 263 ], "score": 1.0, "content": "Data Mining (KDD-98, pp. 73–79. AAAI Press, 1998.", "type": "text" } ], "index": 12 } ], "index": 11 }, { "type": "text", "bbox": [ 104, 270, 504, 293 ], "lines": [ { "bbox": [ 105, 269, 505, 283 ], "spans": [ { "bbox": [ 105, 269, 505, 283 ], "score": 1.0, "content": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 281, 478, 294 ], "spans": [ { "bbox": [ 116, 281, 478, 294 ], "score": 1.0, "content": "In Proceedings of International Conference on Computer Vision (ICCV), December 2015.", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 300, 505, 357 ], "lines": [ { "bbox": [ 106, 301, 505, 313 ], "spans": [ { "bbox": [ 106, 301, 505, 313 ], "score": 1.0, "content": "Mario Luciˇ c, Michael Tschannen, Marvin Ritter, Xiaohua Zhai, Olivier Bachem, and Sylvain Gelly. ´", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 312, 504, 324 ], "spans": [ { "bbox": [ 116, 312, 504, 324 ], "score": 1.0, "content": "High-fidelity image generation with fewer labels. In Kamalika Chaudhuri and Ruslan Salakhut-", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 322, 504, 335 ], "spans": [ { "bbox": [ 116, 322, 504, 335 ], "score": 1.0, "content": "dinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 334, 505, 346 ], "spans": [ { "bbox": [ 116, 334, 505, 346 ], "score": 1.0, "content": "of Proceedings of Machine Learning Research, pp. 4183–4192. PMLR, 09–15 Jun 2019. URL", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 345, 402, 357 ], "spans": [ { "bbox": [ 115, 345, 402, 357 ], "score": 1.0, "content": "https://proceedings.mlr.press/v97/lucic19a.html.", "type": "text" } ], "index": 19 } ], "index": 17 }, { "type": "text", "bbox": [ 107, 364, 504, 388 ], "lines": [ { "bbox": [ 106, 365, 505, 378 ], "spans": [ { "bbox": [ 106, 365, 505, 378 ], "score": 1.0, "content": "Liqian Ma, Xu Jia, Stamatios Georgoulis, Tinne Tuytelaars, and Luc Van Gool. Exemplar guided", "type": "text" } ], "index": 20 }, { "bbox": [ 116, 375, 454, 388 ], "spans": [ { "bbox": [ 116, 375, 454, 388 ], "score": 1.0, "content": "unsupervised image-to-image translation with semantic consistency. In ICLR, 2019.", "type": "text" } ], "index": 21 } ], "index": 20.5 }, { "type": "text", "bbox": [ 107, 395, 504, 430 ], "lines": [ { "bbox": [ 105, 395, 505, 409 ], "spans": [ { "bbox": [ 105, 395, 505, 409 ], "score": 1.0, "content": "Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking genera-", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 406, 506, 420 ], "spans": [ { "bbox": [ 115, 406, 506, 420 ], "score": 1.0, "content": "tive adversarial networks for diverse image synthesis. 2019 IEEE/CVF Conference on Computer", "type": "text" } ], "index": 23 }, { "bbox": [ 117, 418, 374, 430 ], "spans": [ { "bbox": [ 117, 418, 374, 430 ], "score": 1.0, "content": "Vision and Pattern Recognition (CVPR), pp. 1429–1437, 2019a.", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "text", "bbox": [ 104, 437, 504, 461 ], "lines": [ { "bbox": [ 106, 436, 504, 451 ], "spans": [ { "bbox": [ 106, 436, 504, 451 ], "score": 1.0, "content": "Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking genera-", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 448, 403, 462 ], "spans": [ { "bbox": [ 115, 448, 403, 462 ], "score": 1.0, "content": "tive adversarial networks for diverse image synthesis. In CVPR, 2019b.", "type": "text" } ], "index": 26 } ], "index": 25.5 }, { "type": "text", "bbox": [ 107, 468, 504, 502 ], "lines": [ { "bbox": [ 106, 469, 505, 481 ], "spans": [ { "bbox": [ 106, 469, 505, 481 ], "score": 1.0, "content": "Ron Mokady, Sagie Benaim, Lior Wolf, and Amit Bermano. Masked based unsupervised content", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 480, 504, 492 ], "spans": [ { "bbox": [ 116, 480, 504, 492 ], "score": 1.0, "content": "transfer. In International Conference on Learning Representations, 2020. URL https://", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 491, 325, 502 ], "spans": [ { "bbox": [ 116, 491, 254, 502 ], "score": 1.0, "content": "openreview.net/forum?id", "type": "text" }, { "bbox": [ 255, 492, 261, 500 ], "score": 0.56, "content": "=", "type": "inline_equation" }, { "bbox": [ 261, 491, 325, 502 ], "score": 1.0, "content": "BJe-91BtvH.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 510, 505, 555 ], "lines": [ { "bbox": [ 106, 510, 505, 523 ], "spans": [ { "bbox": [ 106, 510, 505, 523 ], "score": 1.0, "content": "Henry Y. T. Ngan, Grantham K. H. Pang, and Nelson H. C. Yung. Review article: Automated", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 521, 505, 534 ], "spans": [ { "bbox": [ 115, 521, 505, 534 ], "score": 1.0, "content": "fabric defect detection-a review. Image Vision Comput., 29(7):442–458, June 2011. ISSN 0262-", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 532, 504, 546 ], "spans": [ { "bbox": [ 115, 532, 504, 546 ], "score": 1.0, "content": "8856. doi: 10.1016/j.imavis.2011.02.002. URL https://doi.org/10.1016/j.imavis.", "type": "text" } ], "index": 32 }, { "bbox": [ 116, 545, 187, 555 ], "spans": [ { "bbox": [ 116, 545, 187, 555 ], "score": 1.0, "content": "2011.02.002.", "type": "text" } ], "index": 33 } ], "index": 31.5 }, { "type": "text", "bbox": [ 106, 563, 505, 597 ], "lines": [ { "bbox": [ 106, 563, 505, 577 ], "spans": [ { "bbox": [ 106, 563, 505, 577 ], "score": 1.0, "content": "Shuanlong Niu, Bin Li, Xinggang Wang, and Hui Lin. Defect image sample generation with gan", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 574, 505, 588 ], "spans": [ { "bbox": [ 115, 574, 505, 588 ], "score": 1.0, "content": "for improving defect recognition. IEEE Transactions on Automation Science and Engineering,", "type": "text" } ], "index": 35 }, { "bbox": [ 117, 586, 359, 598 ], "spans": [ { "bbox": [ 117, 586, 359, 598 ], "score": 1.0, "content": "17(3):1611–1622, 2020. doi: 10.1109/TASE.2020.2967415.", "type": "text" } ], "index": 36 } ], "index": 35 }, { "type": "text", "bbox": [ 104, 605, 505, 628 ], "lines": [ { "bbox": [ 106, 604, 505, 619 ], "spans": [ { "bbox": [ 106, 604, 505, 619 ], "score": 1.0, "content": "Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 617, 317, 628 ], "spans": [ { "bbox": [ 115, 617, 317, 628 ], "score": 1.0, "content": "spatially-adaptive normalization. In CVPR, 2019.", "type": "text" } ], "index": 38 } ], "index": 37.5 }, { "type": "text", "bbox": [ 105, 636, 504, 660 ], "lines": [ { "bbox": [ 105, 634, 505, 650 ], "spans": [ { "bbox": [ 105, 634, 505, 650 ], "score": 1.0, "content": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zach DeVito, Zeming", "type": "text" } ], "index": 39 }, { "bbox": [ 116, 647, 505, 660 ], "spans": [ { "bbox": [ 116, 647, 505, 660 ], "score": 1.0, "content": "Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017.", "type": "text" } ], "index": 40 } ], "index": 39.5 }, { "type": "text", "bbox": [ 106, 667, 505, 701 ], "lines": [ { "bbox": [ 106, 667, 506, 681 ], "spans": [ { "bbox": [ 106, 667, 506, 681 ], "score": 1.0, "content": "Deepak Pathak, Philipp Krahenb ¨ uhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context ¨", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 678, 505, 691 ], "spans": [ { "bbox": [ 115, 678, 505, 691 ], "score": 1.0, "content": "encoders: Feature learning by inpainting. 2016 IEEE Conference on Computer Vision and Pattern", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 689, 293, 702 ], "spans": [ { "bbox": [ 116, 689, 293, 702 ], "score": 1.0, "content": "Recognition (CVPR), pp. 2536–2544, 2016.", "type": "text" } ], "index": 43 } ], "index": 42 }, { "type": "text", "bbox": [ 106, 709, 504, 732 ], "lines": [ { "bbox": [ 105, 708, 506, 723 ], "spans": [ { "bbox": [ 105, 708, 506, 723 ], "score": 1.0, "content": "Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv preprint", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 720, 220, 732 ], "spans": [ { "bbox": [ 115, 720, 220, 732 ], "score": 1.0, "content": "arXiv:1609.04747, 2016.", "type": "text" } ], "index": 45 } ], "index": 44.5 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 505, 105 ], "lines": [ { "bbox": [ 105, 79, 506, 96 ], "spans": [ { "bbox": [ 105, 79, 506, 96 ], "score": 1.0, "content": "Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 347, 105 ], "spans": [ { "bbox": [ 115, 93, 347, 105 ], "score": 1.0, "content": "generative adversarial networks with limited data, 2020a.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 79, 506, 105 ] }, { "type": "text", "bbox": [ 108, 113, 503, 147 ], "lines": [ { "bbox": [ 105, 112, 505, 127 ], "spans": [ { "bbox": [ 105, 112, 505, 127 ], "score": 1.0, "content": "Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyz-", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 125, 505, 137 ], "spans": [ { "bbox": [ 117, 125, 505, 137 ], "score": 1.0, "content": "ing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 135, 373, 149 ], "spans": [ { "bbox": [ 117, 135, 373, 149 ], "score": 1.0, "content": "Computer Vision and Pattern Recognition (CVPR), June 2020b.", "type": "text" } ], "index": 4 } ], "index": 3, "bbox_fs": [ 105, 112, 505, 149 ] }, { "type": "text", "bbox": [ 106, 155, 503, 178 ], "lines": [ { "bbox": [ 105, 154, 505, 169 ], "spans": [ { "bbox": [ 105, 154, 505, 169 ], "score": 1.0, "content": "Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Tech-", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 167, 366, 178 ], "spans": [ { "bbox": [ 115, 167, 366, 178 ], "score": 1.0, "content": "nical Report 0, University of Toronto, Toronto, Ontario, 2009.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 105, 154, 505, 178 ] }, { "type": "text", "bbox": [ 106, 186, 506, 220 ], "lines": [ { "bbox": [ 106, 186, 505, 199 ], "spans": [ { "bbox": [ 106, 186, 505, 199 ], "score": 1.0, "content": "Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro ´", "type": "text" } ], "index": 7 }, { "bbox": [ 116, 198, 505, 210 ], "spans": [ { "bbox": [ 116, 198, 505, 210 ], "score": 1.0, "content": "Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photo-realistic single", "type": "text" } ], "index": 8 }, { "bbox": [ 116, 209, 436, 221 ], "spans": [ { "bbox": [ 116, 209, 436, 221 ], "score": 1.0, "content": "image super-resolution using a generative adversarial network. In CVPR, 2017.", "type": "text" } ], "index": 9 } ], "index": 8, "bbox_fs": [ 106, 186, 505, 221 ] }, { "type": "text", "bbox": [ 106, 227, 506, 262 ], "lines": [ { "bbox": [ 106, 228, 505, 240 ], "spans": [ { "bbox": [ 106, 228, 505, 240 ], "score": 1.0, "content": "Charles Ling, , Charles X. Ling, and Chenghui Li. Data mining for direct marketing: Problems and", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 240, 506, 252 ], "spans": [ { "bbox": [ 116, 240, 506, 252 ], "score": 1.0, "content": "solutions. In In Proceedings of the Fourth International Conference on Knowledge Discovery and", "type": "text" } ], "index": 11 }, { "bbox": [ 116, 250, 336, 263 ], "spans": [ { "bbox": [ 116, 250, 336, 263 ], "score": 1.0, "content": "Data Mining (KDD-98, pp. 73–79. AAAI Press, 1998.", "type": "text" } ], "index": 12 } ], "index": 11, "bbox_fs": [ 106, 228, 506, 263 ] }, { "type": "text", "bbox": [ 104, 270, 504, 293 ], "lines": [ { "bbox": [ 105, 269, 505, 283 ], "spans": [ { "bbox": [ 105, 269, 505, 283 ], "score": 1.0, "content": "Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.", "type": "text" } ], "index": 13 }, { "bbox": [ 116, 281, 478, 294 ], "spans": [ { "bbox": [ 116, 281, 478, 294 ], "score": 1.0, "content": "In Proceedings of International Conference on Computer Vision (ICCV), December 2015.", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 105, 269, 505, 294 ] }, { "type": "text", "bbox": [ 106, 300, 505, 357 ], "lines": [ { "bbox": [ 106, 301, 505, 313 ], "spans": [ { "bbox": [ 106, 301, 505, 313 ], "score": 1.0, "content": "Mario Luciˇ c, Michael Tschannen, Marvin Ritter, Xiaohua Zhai, Olivier Bachem, and Sylvain Gelly. ´", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 312, 504, 324 ], "spans": [ { "bbox": [ 116, 312, 504, 324 ], "score": 1.0, "content": "High-fidelity image generation with fewer labels. In Kamalika Chaudhuri and Ruslan Salakhut-", "type": "text" } ], "index": 16 }, { "bbox": [ 116, 322, 504, 335 ], "spans": [ { "bbox": [ 116, 322, 504, 335 ], "score": 1.0, "content": "dinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 334, 505, 346 ], "spans": [ { "bbox": [ 116, 334, 505, 346 ], "score": 1.0, "content": "of Proceedings of Machine Learning Research, pp. 4183–4192. PMLR, 09–15 Jun 2019. URL", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 345, 402, 357 ], "spans": [ { "bbox": [ 115, 345, 402, 357 ], "score": 1.0, "content": "https://proceedings.mlr.press/v97/lucic19a.html.", "type": "text" } ], "index": 19 } ], "index": 17, "bbox_fs": [ 106, 301, 505, 357 ] }, { "type": "text", "bbox": [ 107, 364, 504, 388 ], "lines": [ { "bbox": [ 106, 365, 505, 378 ], "spans": [ { "bbox": [ 106, 365, 505, 378 ], "score": 1.0, "content": "Liqian Ma, Xu Jia, Stamatios Georgoulis, Tinne Tuytelaars, and Luc Van Gool. Exemplar guided", "type": "text" } ], "index": 20 }, { "bbox": [ 116, 375, 454, 388 ], "spans": [ { "bbox": [ 116, 375, 454, 388 ], "score": 1.0, "content": "unsupervised image-to-image translation with semantic consistency. In ICLR, 2019.", "type": "text" } ], "index": 21 } ], "index": 20.5, "bbox_fs": [ 106, 365, 505, 388 ] }, { "type": "text", "bbox": [ 107, 395, 504, 430 ], "lines": [ { "bbox": [ 105, 395, 505, 409 ], "spans": [ { "bbox": [ 105, 395, 505, 409 ], "score": 1.0, "content": "Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking genera-", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 406, 506, 420 ], "spans": [ { "bbox": [ 115, 406, 506, 420 ], "score": 1.0, "content": "tive adversarial networks for diverse image synthesis. 2019 IEEE/CVF Conference on Computer", "type": "text" } ], "index": 23 }, { "bbox": [ 117, 418, 374, 430 ], "spans": [ { "bbox": [ 117, 418, 374, 430 ], "score": 1.0, "content": "Vision and Pattern Recognition (CVPR), pp. 1429–1437, 2019a.", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 105, 395, 506, 430 ] }, { "type": "text", "bbox": [ 104, 437, 504, 461 ], "lines": [ { "bbox": [ 106, 436, 504, 451 ], "spans": [ { "bbox": [ 106, 436, 504, 451 ], "score": 1.0, "content": "Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking genera-", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 448, 403, 462 ], "spans": [ { "bbox": [ 115, 448, 403, 462 ], "score": 1.0, "content": "tive adversarial networks for diverse image synthesis. In CVPR, 2019b.", "type": "text" } ], "index": 26 } ], "index": 25.5, "bbox_fs": [ 106, 436, 504, 462 ] }, { "type": "text", "bbox": [ 107, 468, 504, 502 ], "lines": [ { "bbox": [ 106, 469, 505, 481 ], "spans": [ { "bbox": [ 106, 469, 505, 481 ], "score": 1.0, "content": "Ron Mokady, Sagie Benaim, Lior Wolf, and Amit Bermano. Masked based unsupervised content", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 480, 504, 492 ], "spans": [ { "bbox": [ 116, 480, 504, 492 ], "score": 1.0, "content": "transfer. In International Conference on Learning Representations, 2020. URL https://", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 491, 325, 502 ], "spans": [ { "bbox": [ 116, 491, 254, 502 ], "score": 1.0, "content": "openreview.net/forum?id", "type": "text" }, { "bbox": [ 255, 492, 261, 500 ], "score": 0.56, "content": "=", "type": "inline_equation" }, { "bbox": [ 261, 491, 325, 502 ], "score": 1.0, "content": "BJe-91BtvH.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 106, 469, 505, 502 ] }, { "type": "text", "bbox": [ 106, 510, 505, 555 ], "lines": [ { "bbox": [ 106, 510, 505, 523 ], "spans": [ { "bbox": [ 106, 510, 505, 523 ], "score": 1.0, "content": "Henry Y. T. Ngan, Grantham K. H. Pang, and Nelson H. C. Yung. Review article: Automated", "type": "text" } ], "index": 30 }, { "bbox": [ 115, 521, 505, 534 ], "spans": [ { "bbox": [ 115, 521, 505, 534 ], "score": 1.0, "content": "fabric defect detection-a review. Image Vision Comput., 29(7):442–458, June 2011. ISSN 0262-", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 532, 504, 546 ], "spans": [ { "bbox": [ 115, 532, 504, 546 ], "score": 1.0, "content": "8856. doi: 10.1016/j.imavis.2011.02.002. URL https://doi.org/10.1016/j.imavis.", "type": "text" } ], "index": 32 }, { "bbox": [ 116, 545, 187, 555 ], "spans": [ { "bbox": [ 116, 545, 187, 555 ], "score": 1.0, "content": "2011.02.002.", "type": "text" } ], "index": 33 } ], "index": 31.5, "bbox_fs": [ 106, 510, 505, 555 ] }, { "type": "text", "bbox": [ 106, 563, 505, 597 ], "lines": [ { "bbox": [ 106, 563, 505, 577 ], "spans": [ { "bbox": [ 106, 563, 505, 577 ], "score": 1.0, "content": "Shuanlong Niu, Bin Li, Xinggang Wang, and Hui Lin. Defect image sample generation with gan", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 574, 505, 588 ], "spans": [ { "bbox": [ 115, 574, 505, 588 ], "score": 1.0, "content": "for improving defect recognition. IEEE Transactions on Automation Science and Engineering,", "type": "text" } ], "index": 35 }, { "bbox": [ 117, 586, 359, 598 ], "spans": [ { "bbox": [ 117, 586, 359, 598 ], "score": 1.0, "content": "17(3):1611–1622, 2020. doi: 10.1109/TASE.2020.2967415.", "type": "text" } ], "index": 36 } ], "index": 35, "bbox_fs": [ 106, 563, 505, 598 ] }, { "type": "text", "bbox": [ 104, 605, 505, 628 ], "lines": [ { "bbox": [ 106, 604, 505, 619 ], "spans": [ { "bbox": [ 106, 604, 505, 619 ], "score": 1.0, "content": "Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 617, 317, 628 ], "spans": [ { "bbox": [ 115, 617, 317, 628 ], "score": 1.0, "content": "spatially-adaptive normalization. In CVPR, 2019.", "type": "text" } ], "index": 38 } ], "index": 37.5, "bbox_fs": [ 106, 604, 505, 628 ] }, { "type": "text", "bbox": [ 105, 636, 504, 660 ], "lines": [ { "bbox": [ 105, 634, 505, 650 ], "spans": [ { "bbox": [ 105, 634, 505, 650 ], "score": 1.0, "content": "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zach DeVito, Zeming", "type": "text" } ], "index": 39 }, { "bbox": [ 116, 647, 505, 660 ], "spans": [ { "bbox": [ 116, 647, 505, 660 ], "score": 1.0, "content": "Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017.", "type": "text" } ], "index": 40 } ], "index": 39.5, "bbox_fs": [ 105, 634, 505, 660 ] }, { "type": "text", "bbox": [ 106, 667, 505, 701 ], "lines": [ { "bbox": [ 106, 667, 506, 681 ], "spans": [ { "bbox": [ 106, 667, 506, 681 ], "score": 1.0, "content": "Deepak Pathak, Philipp Krahenb ¨ uhl, Jeff Donahue, Trevor Darrell, and Alexei A. Efros. Context ¨", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 678, 505, 691 ], "spans": [ { "bbox": [ 115, 678, 505, 691 ], "score": 1.0, "content": "encoders: Feature learning by inpainting. 2016 IEEE Conference on Computer Vision and Pattern", "type": "text" } ], "index": 42 }, { "bbox": [ 116, 689, 293, 702 ], "spans": [ { "bbox": [ 116, 689, 293, 702 ], "score": 1.0, "content": "Recognition (CVPR), pp. 2536–2544, 2016.", "type": "text" } ], "index": 43 } ], "index": 42, "bbox_fs": [ 106, 667, 506, 702 ] }, { "type": "text", "bbox": [ 106, 709, 504, 732 ], "lines": [ { "bbox": [ 105, 708, 506, 723 ], "spans": [ { "bbox": [ 105, 708, 506, 723 ], "score": 1.0, "content": "Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv preprint", "type": "text" } ], "index": 44 }, { "bbox": [ 115, 720, 220, 732 ], "spans": [ { "bbox": [ 115, 720, 220, 732 ], "score": 1.0, "content": "arXiv:1609.04747, 2016.", "type": "text" } ], "index": 45 } ], "index": 44.5, "bbox_fs": [ 105, 708, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 503, 105 ], "lines": [ { "bbox": [ 105, 81, 505, 96 ], "spans": [ { "bbox": [ 105, 81, 505, 96 ], "score": 1.0, "content": "Connor Shorten and T. Khoshgoftaar. A survey on image data augmentation for deep learning.", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 94, 259, 105 ], "spans": [ { "bbox": [ 116, 94, 259, 105 ], "score": 1.0, "content": "Journal of Big Data, 6:1–48, 2019.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 107, 112, 504, 168 ], "lines": [ { "bbox": [ 106, 113, 505, 124 ], "spans": [ { "bbox": [ 106, 113, 505, 124 ], "score": 1.0, "content": "Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural net-", "type": "text" } ], "index": 2 }, { "bbox": [ 116, 123, 505, 136 ], "spans": [ { "bbox": [ 116, 123, 505, 136 ], "score": 1.0, "content": "works. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th In-", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 133, 506, 149 ], "spans": [ { "bbox": [ 115, 133, 506, 149 ], "score": 1.0, "content": "ternational Conference on Machine Learning, volume 97 of Proceedings of Machine Learning", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 143, 505, 159 ], "spans": [ { "bbox": [ 115, 143, 505, 159 ], "score": 1.0, "content": "Research, pp. 6105–6114. PMLR, 09–15 Jun 2019. URL https://proceedings.mlr.", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 156, 247, 168 ], "spans": [ { "bbox": [ 115, 156, 247, 168 ], "score": 1.0, "content": "press/v97/tan19a.html.", "type": "text" } ], "index": 6 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 174, 503, 198 ], "lines": [ { "bbox": [ 105, 173, 505, 188 ], "spans": [ { "bbox": [ 105, 173, 505, 188 ], "score": 1.0, "content": "Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tiangchen Zhao, and Honglak Lee. Diversity-", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 186, 395, 198 ], "spans": [ { "bbox": [ 115, 186, 395, 198 ], "score": 1.0, "content": "sensitive conditional generative adversarial networks. In ICLR, 2019.", "type": "text" } ], "index": 8 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 204, 505, 239 ], "lines": [ { "bbox": [ 105, 203, 505, 219 ], "spans": [ { "bbox": [ 105, 203, 505, 219 ], "score": 1.0, "content": "Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-gan: High-fidelity defect syn-", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 216, 505, 228 ], "spans": [ { "bbox": [ 116, 216, 505, 228 ], "score": 1.0, "content": "thesis for automated defect inspection. In Proceedings of the IEEE/CVF Winter Conference on", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 227, 411, 240 ], "spans": [ { "bbox": [ 115, 227, 411, 240 ], "score": 1.0, "content": "Applications of Computer Vision (WACV), pp. 2524–2534, January 2021.", "type": "text" } ], "index": 11 } ], "index": 10 }, { "type": "text", "bbox": [ 106, 245, 504, 268 ], "lines": [ { "bbox": [ 105, 245, 506, 258 ], "spans": [ { "bbox": [ 105, 245, 506, 258 ], "score": 1.0, "content": "Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 256, 249, 269 ], "spans": [ { "bbox": [ 116, 256, 249, 269 ], "score": 1.0, "content": "data-efficient gan training, 2020.", "type": "text" } ], "index": 13 } ], "index": 12.5 }, { "type": "text", "bbox": [ 106, 275, 505, 309 ], "lines": [ { "bbox": [ 105, 275, 505, 289 ], "spans": [ { "bbox": [ 105, 275, 505, 289 ], "score": 1.0, "content": "Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation", "type": "text" } ], "index": 14 }, { "bbox": [ 116, 287, 505, 299 ], "spans": [ { "bbox": [ 116, 287, 505, 299 ], "score": 1.0, "content": "using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 298, 279, 310 ], "spans": [ { "bbox": [ 116, 298, 279, 310 ], "score": 1.0, "content": "on Computer Vision (ICCV), Oct 2017a.", "type": "text" } ], "index": 16 } ], "index": 15 }, { "type": "text", "bbox": [ 106, 316, 504, 339 ], "lines": [ { "bbox": [ 106, 316, 505, 329 ], "spans": [ { "bbox": [ 106, 316, 505, 329 ], "score": 1.0, "content": "Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A. Efros, O. Wang, and", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 328, 440, 339 ], "spans": [ { "bbox": [ 116, 328, 440, 339 ], "score": 1.0, "content": "E. Shechtman. Toward multimodal image-to-image translation. In NIPS, 2017b.", "type": "text" } ], "index": 18 } ], "index": 17.5 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "12", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 503, 105 ], "lines": [ { "bbox": [ 105, 81, 505, 96 ], "spans": [ { "bbox": [ 105, 81, 505, 96 ], "score": 1.0, "content": "Connor Shorten and T. Khoshgoftaar. A survey on image data augmentation for deep learning.", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 94, 259, 105 ], "spans": [ { "bbox": [ 116, 94, 259, 105 ], "score": 1.0, "content": "Journal of Big Data, 6:1–48, 2019.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 81, 505, 105 ] }, { "type": "text", "bbox": [ 107, 112, 504, 168 ], "lines": [ { "bbox": [ 106, 113, 505, 124 ], "spans": [ { "bbox": [ 106, 113, 505, 124 ], "score": 1.0, "content": "Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural net-", "type": "text" } ], "index": 2 }, { "bbox": [ 116, 123, 505, 136 ], "spans": [ { "bbox": [ 116, 123, 505, 136 ], "score": 1.0, "content": "works. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th In-", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 133, 506, 149 ], "spans": [ { "bbox": [ 115, 133, 506, 149 ], "score": 1.0, "content": "ternational Conference on Machine Learning, volume 97 of Proceedings of Machine Learning", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 143, 505, 159 ], "spans": [ { "bbox": [ 115, 143, 505, 159 ], "score": 1.0, "content": "Research, pp. 6105–6114. PMLR, 09–15 Jun 2019. URL https://proceedings.mlr.", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 156, 247, 168 ], "spans": [ { "bbox": [ 115, 156, 247, 168 ], "score": 1.0, "content": "press/v97/tan19a.html.", "type": "text" } ], "index": 6 } ], "index": 4, "bbox_fs": [ 106, 113, 506, 168 ] }, { "type": "text", "bbox": [ 107, 174, 503, 198 ], "lines": [ { "bbox": [ 105, 173, 505, 188 ], "spans": [ { "bbox": [ 105, 173, 505, 188 ], "score": 1.0, "content": "Dingdong Yang, Seunghoon Hong, Yunseok Jang, Tiangchen Zhao, and Honglak Lee. Diversity-", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 186, 395, 198 ], "spans": [ { "bbox": [ 115, 186, 395, 198 ], "score": 1.0, "content": "sensitive conditional generative adversarial networks. In ICLR, 2019.", "type": "text" } ], "index": 8 } ], "index": 7.5, "bbox_fs": [ 105, 173, 505, 198 ] }, { "type": "text", "bbox": [ 107, 204, 505, 239 ], "lines": [ { "bbox": [ 105, 203, 505, 219 ], "spans": [ { "bbox": [ 105, 203, 505, 219 ], "score": 1.0, "content": "Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-gan: High-fidelity defect syn-", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 216, 505, 228 ], "spans": [ { "bbox": [ 116, 216, 505, 228 ], "score": 1.0, "content": "thesis for automated defect inspection. In Proceedings of the IEEE/CVF Winter Conference on", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 227, 411, 240 ], "spans": [ { "bbox": [ 115, 227, 411, 240 ], "score": 1.0, "content": "Applications of Computer Vision (WACV), pp. 2524–2534, January 2021.", "type": "text" } ], "index": 11 } ], "index": 10, "bbox_fs": [ 105, 203, 505, 240 ] }, { "type": "text", "bbox": [ 106, 245, 504, 268 ], "lines": [ { "bbox": [ 105, 245, 506, 258 ], "spans": [ { "bbox": [ 105, 245, 506, 258 ], "score": 1.0, "content": "Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 256, 249, 269 ], "spans": [ { "bbox": [ 116, 256, 249, 269 ], "score": 1.0, "content": "data-efficient gan training, 2020.", "type": "text" } ], "index": 13 } ], "index": 12.5, "bbox_fs": [ 105, 245, 506, 269 ] }, { "type": "text", "bbox": [ 106, 275, 505, 309 ], "lines": [ { "bbox": [ 105, 275, 505, 289 ], "spans": [ { "bbox": [ 105, 275, 505, 289 ], "score": 1.0, "content": "Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation", "type": "text" } ], "index": 14 }, { "bbox": [ 116, 287, 505, 299 ], "spans": [ { "bbox": [ 116, 287, 505, 299 ], "score": 1.0, "content": "using cycle-consistent adversarial networks. In Proceedings of the IEEE International Conference", "type": "text" } ], "index": 15 }, { "bbox": [ 116, 298, 279, 310 ], "spans": [ { "bbox": [ 116, 298, 279, 310 ], "score": 1.0, "content": "on Computer Vision (ICCV), Oct 2017a.", "type": "text" } ], "index": 16 } ], "index": 15, "bbox_fs": [ 105, 275, 505, 310 ] }, { "type": "text", "bbox": [ 106, 316, 504, 339 ], "lines": [ { "bbox": [ 106, 316, 505, 329 ], "spans": [ { "bbox": [ 106, 316, 505, 329 ], "score": 1.0, "content": "Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A. Efros, O. Wang, and", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 328, 440, 339 ], "spans": [ { "bbox": [ 116, 328, 440, 339 ], "score": 1.0, "content": "E. Shechtman. Toward multimodal image-to-image translation. In NIPS, 2017b.", "type": "text" } ], "index": 18 } ], "index": 17.5, "bbox_fs": [ 106, 316, 505, 339 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 81, 361, 94 ], "lines": [ { "bbox": [ 106, 81, 362, 95 ], "spans": [ { "bbox": [ 106, 81, 362, 95 ], "score": 1.0, "content": "A THE SURFACE DEFECT INSPECTION DATASET", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 505, 183 ], "lines": [ { "bbox": [ 106, 106, 505, 118 ], "spans": [ { "bbox": [ 106, 106, 412, 118 ], "score": 1.0, "content": "The Surface Defect Inspection (SDI) dataset consists of 20,414 images at", "type": "text" }, { "bbox": [ 413, 106, 458, 117 ], "score": 0.89, "content": "1 2 8 \\times 1 2 8", "type": "inline_equation" }, { "bbox": [ 459, 106, 505, 118 ], "score": 1.0, "content": "resolution.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 117, 505, 129 ], "spans": [ { "bbox": [ 105, 117, 349, 129 ], "score": 1.0, "content": "It contains three background domains—product A, product", "type": "text" }, { "bbox": [ 349, 118, 358, 127 ], "score": 0.3, "content": "\\mathbf { B }", "type": "inline_equation" }, { "bbox": [ 358, 117, 505, 129 ], "score": 1.0, "content": "and product C, each can be further", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 128, 505, 140 ], "spans": [ { "bbox": [ 105, 128, 505, 140 ], "score": 1.0, "content": "classified into three foreground domains—Normal, Scratches and Spots. Figure 7 shows", "type": "text" } ], "index": 3 }, { "bbox": [ 104, 138, 505, 153 ], "spans": [ { "bbox": [ 104, 138, 505, 153 ], "score": 1.0, "content": "example images of the SDI dataset. To be noticed that the dataset is highly imbalanced not only", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 150, 504, 161 ], "spans": [ { "bbox": [ 106, 150, 504, 161 ], "score": 1.0, "content": "between normal and defective samples but also between different products as shown in Table 5.", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 160, 506, 174 ], "spans": [ { "bbox": [ 105, 160, 506, 174 ], "score": 1.0, "content": "This sets a more challenging task when training deep neural networks like GANs and downstream", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 171, 151, 184 ], "spans": [ { "bbox": [ 105, 171, 151, 184 ], "score": 1.0, "content": "classifiers.", "type": "text" } ], "index": 7 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 188, 505, 288 ], "lines": [ { "bbox": [ 105, 187, 505, 201 ], "spans": [ { "bbox": [ 105, 187, 505, 201 ], "score": 1.0, "content": "For each foreground and background domains, we randomly select 50 images for a joint valida-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 199, 505, 213 ], "spans": [ { "bbox": [ 105, 199, 505, 213 ], "score": 1.0, "content": "tion/test set, which is then further split into separate sets in the ratio of 3:7, and use all remaining", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 210, 506, 225 ], "spans": [ { "bbox": [ 105, 210, 506, 225 ], "score": 1.0, "content": "images as training sets for GAN and classifier training. We present the distribution of the training", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 222, 504, 235 ], "spans": [ { "bbox": [ 105, 222, 504, 235 ], "score": 1.0, "content": "set when training DT-GAN in Table 6. Note that the normal samples used in GAN training are only", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 232, 506, 245 ], "spans": [ { "bbox": [ 105, 232, 506, 245 ], "score": 1.0, "content": "a subset of all available samples in Normal and we keep the rest of them for generating defective", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 243, 505, 256 ], "spans": [ { "bbox": [ 105, 243, 505, 256 ], "score": 1.0, "content": "samples at test time. For classifier training, we show the statistics in Table 7, where the number", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 254, 506, 267 ], "spans": [ { "bbox": [ 105, 254, 506, 267 ], "score": 1.0, "content": "of normal samples involved in classifier training increase incrementally. The validation set is used", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 265, 505, 277 ], "spans": [ { "bbox": [ 105, 265, 505, 277 ], "score": 1.0, "content": "to select the best model during classifier training while the test set is left untouched until the final", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 277, 409, 289 ], "spans": [ { "bbox": [ 105, 277, 409, 289 ], "score": 1.0, "content": "evaluation. Both of the validation and test set are inaccessible by DT-GAN.", "type": "text" } ], "index": 16 } ], "index": 12 }, { "type": "table", "bbox": [ 118, 325, 297, 393 ], "blocks": [ { "type": "table_caption", "bbox": [ 120, 314, 295, 324 ], "group_id": 0, "lines": [ { "bbox": [ 118, 311, 297, 326 ], "spans": [ { "bbox": [ 118, 311, 297, 326 ], "score": 1.0, "content": "Table 5: Distribution of the full SDI dataset.", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "table_body", "bbox": [ 118, 325, 297, 393 ], "group_id": 0, "lines": [ { "bbox": [ 118, 325, 297, 393 ], "spans": [ { "bbox": [ 118, 325, 297, 393 ], "score": 0.979, "html": "
Overview
ABC
Normal625062506250
Scratches340167121
Spots108670258
", "type": "table", "image_path": "c5ad4cce383f42b999dbec5baa09abd0eec003f8449b17cca61eb4e7218407fe.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 118, 325, 297, 338.6 ], "spans": [], "index": 18 }, { "bbox": [ 118, 338.6, 297, 352.20000000000005 ], "spans": [], "index": 19 }, { "bbox": [ 118, 352.20000000000005, 297, 365.80000000000007 ], "spans": [], "index": 20 }, { "bbox": [ 118, 365.80000000000007, 297, 379.4000000000001 ], "spans": [], "index": 21 }, { "bbox": [ 118, 379.4000000000001, 297, 393.0000000000001 ], "spans": [], "index": 22 } ] } ], "index": 18.5 }, { "type": "table", "bbox": [ 315, 330, 495, 399 ], "blocks": [ { "type": "table_caption", "bbox": [ 314, 307, 495, 329 ], "group_id": 2, "lines": [ { "bbox": [ 315, 307, 496, 319 ], "spans": [ { "bbox": [ 315, 307, 496, 319 ], "score": 1.0, "content": "Table 6: The training set for DT-GAN and", "type": "text" } ], "index": 23 }, { "bbox": [ 314, 318, 469, 331 ], "spans": [ { "bbox": [ 314, 318, 469, 331 ], "score": 1.0, "content": "the baseline image synthesis methods.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "table_body", "bbox": [ 315, 330, 495, 399 ], "group_id": 2, "lines": [ { "bbox": [ 315, 330, 495, 399 ], "spans": [ { "bbox": [ 315, 330, 495, 399 ], "score": 0.977, "html": "
Overview
ABC
Normal700700700
Scratches29011771
Spots58620208
", "type": "table", "image_path": "d3b1a834deec1f06f6598262b6c7958b3d090165d454d73f8326527e6462444b.jpg" } ] } ], "index": 27, "virtual_lines": [ { "bbox": [ 315, 330, 495, 343.8 ], "spans": [], "index": 25 }, { "bbox": [ 315, 343.8, 495, 357.6 ], "spans": [], "index": 26 }, { "bbox": [ 315, 357.6, 495, 371.40000000000003 ], "spans": [], "index": 27 }, { "bbox": [ 315, 371.40000000000003, 495, 385.20000000000005 ], "spans": [], "index": 28 }, { "bbox": [ 315, 385.20000000000005, 495, 399.00000000000006 ], "spans": [], "index": 29 } ] } ], "index": 25.25 }, { "type": "table", "bbox": [ 106, 447, 510, 517 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 425, 505, 446 ], "group_id": 1, "lines": [ { "bbox": [ 105, 423, 505, 438 ], "spans": [ { "bbox": [ 105, 423, 450, 438 ], "score": 1.0, "content": "Table 7: The training, validation and test set for classifier training, where", "type": "text" }, { "bbox": [ 451, 425, 461, 435 ], "score": 0.71, "content": "N", "type": "inline_equation" }, { "bbox": [ 462, 423, 505, 438 ], "score": 1.0, "content": "increases", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 435, 284, 447 ], "spans": [ { "bbox": [ 106, 435, 284, 447 ], "score": 1.0, "content": "incrementally—1500, 2200, 4000 and 6200.", "type": "text" } ], "index": 31 } ], "index": 30.5 }, { "type": "table_body", "bbox": [ 106, 447, 510, 517 ], "group_id": 1, "lines": [ { "bbox": [ 106, 447, 510, 517 ], "spans": [ { "bbox": [ 106, 447, 510, 517 ], "score": 0.978, "html": "
TrainValidationTest
ABCABCABC
NormalNNN121815383235
Scratches29011771141615363435
Spots58620208141615363435
", "type": "table", "image_path": "ecde65a45fe9d34b0e307da325180a667544009c522094149b40f88f28ea42b7.jpg" } ] } ], "index": 33, "virtual_lines": [ { "bbox": [ 106, 447, 510, 470.3333333333333 ], "spans": [], "index": 32 }, { "bbox": [ 106, 470.3333333333333, 510, 493.66666666666663 ], "spans": [], "index": 33 }, { "bbox": [ 106, 493.66666666666663, 510, 517.0 ], "spans": [], "index": 34 } ] } ], "index": 31.75 }, { "type": "title", "bbox": [ 108, 539, 228, 552 ], "lines": [ { "bbox": [ 105, 538, 230, 554 ], "spans": [ { "bbox": [ 105, 538, 230, 554 ], "score": 1.0, "content": "B TRAINING DETAILS", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 106, 564, 505, 630 ], "lines": [ { "bbox": [ 105, 563, 506, 576 ], "spans": [ { "bbox": [ 105, 563, 506, 576 ], "score": 1.0, "content": "DT-GAN. We follow the training scheme as described in StarGAN v2 with minor modifications. To", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 575, 505, 587 ], "spans": [ { "bbox": [ 105, 575, 505, 587 ], "score": 1.0, "content": "fit the model on a single Nvidia GTX TITAN X, the batch size is reduced to four while the model is", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 586, 506, 599 ], "spans": [ { "bbox": [ 105, 586, 506, 599 ], "score": 1.0, "content": "still trained for 100,000 iterations. The training time is about three and a half days on the dedicated", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 597, 504, 609 ], "spans": [ { "bbox": [ 106, 597, 504, 609 ], "score": 1.0, "content": "GPU with the modified network architecture2 and loss functions mentioned in Section 3 in PyTorch", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 607, 506, 621 ], "spans": [ { "bbox": [ 105, 607, 224, 621 ], "score": 1.0, "content": "(Paszke et al., 2017). We set", "type": "text" }, { "bbox": [ 224, 608, 255, 620 ], "score": 0.89, "content": "\\lambda _ { \\mathrm { s t y } } = 1", "type": "inline_equation" }, { "bbox": [ 256, 607, 259, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 260, 608, 289, 619 ], "score": 0.86, "content": "\\lambda _ { \\mathrm { d s } } = 1", "type": "inline_equation" }, { "bbox": [ 290, 607, 293, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 294, 608, 327, 620 ], "score": 0.88, "content": "\\lambda _ { \\mathrm { c y c } } = 1", "type": "inline_equation" }, { "bbox": [ 327, 607, 331, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 331, 608, 377, 620 ], "score": 0.86, "content": "\\lambda _ { \\mathrm { c o n . c y c } } = 1", "type": "inline_equation" }, { "bbox": [ 377, 607, 380, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 381, 608, 412, 619 ], "score": 0.89, "content": "\\lambda _ { \\mathrm { c l s } } = 1", "type": "inline_equation" }, { "bbox": [ 413, 607, 431, 621 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 431, 608, 474, 619 ], "score": 0.92, "content": "\\lambda _ { \\mathrm { B G . c l s } } = 1", "type": "inline_equation" }, { "bbox": [ 475, 607, 506, 621 ], "score": 1.0, "content": "for the", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 618, 402, 631 ], "spans": [ { "bbox": [ 105, 618, 402, 631 ], "score": 1.0, "content": "SDI dataset. All other design choices remain the same as in StarGAN v2.", "type": "text" } ], "index": 41 } ], "index": 38.5 }, { "type": "text", "bbox": [ 106, 635, 505, 713 ], "lines": [ { "bbox": [ 106, 636, 505, 648 ], "spans": [ { "bbox": [ 106, 636, 505, 648 ], "score": 1.0, "content": "Classifiers. We train all the classifiers that use ResNet-50 as backbone for 100 epochs with the", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 646, 506, 659 ], "spans": [ { "bbox": [ 105, 646, 506, 659 ], "score": 1.0, "content": "SGD optimizer (Ruder, 2016) and batch size 256. The initial learning rate is 0.001, momentum is", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 658, 505, 670 ], "spans": [ { "bbox": [ 105, 658, 505, 670 ], "score": 1.0, "content": "0.9 and weight decay is 1e-4. A learning rate scheduler is set to reduce the learning rate by factor", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 668, 505, 681 ], "spans": [ { "bbox": [ 105, 668, 505, 681 ], "score": 1.0, "content": "of 0.1 when the validation loss stops decreasing for 5 epochs. The same setting also applies to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 679, 505, 692 ], "spans": [ { "bbox": [ 105, 679, 505, 692 ], "score": 1.0, "content": "EfficientNet-b4, except the batch size is reduced to 128. Although DT-GAN can synthesize realistic", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 691, 506, 704 ], "spans": [ { "bbox": [ 105, 691, 506, 704 ], "score": 1.0, "content": "defective samples, we notice that there still exists a domain gap between the generated samples", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 702, 505, 716 ], "spans": [ { "bbox": [ 105, 702, 505, 716 ], "score": 1.0, "content": "and the real samples. To explore the full potential of the generated samples, we attach an auxiliary", "type": "text" } ], "index": 48 } ], "index": 45 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 115, 721, 493, 732 ], "lines": [ { "bbox": [ 118, 718, 494, 734 ], "spans": [ { "bbox": [ 118, 718, 494, 734 ], "score": 1.0, "content": "2We based our implementation on source code from StarGAN v2: https://github.com/clovaai/stargan-v2", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 81, 361, 94 ], "lines": [ { "bbox": [ 106, 81, 362, 95 ], "spans": [ { "bbox": [ 106, 81, 362, 95 ], "score": 1.0, "content": "A THE SURFACE DEFECT INSPECTION DATASET", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 106, 505, 183 ], "lines": [ { "bbox": [ 106, 106, 505, 118 ], "spans": [ { "bbox": [ 106, 106, 412, 118 ], "score": 1.0, "content": "The Surface Defect Inspection (SDI) dataset consists of 20,414 images at", "type": "text" }, { "bbox": [ 413, 106, 458, 117 ], "score": 0.89, "content": "1 2 8 \\times 1 2 8", "type": "inline_equation" }, { "bbox": [ 459, 106, 505, 118 ], "score": 1.0, "content": "resolution.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 117, 505, 129 ], "spans": [ { "bbox": [ 105, 117, 349, 129 ], "score": 1.0, "content": "It contains three background domains—product A, product", "type": "text" }, { "bbox": [ 349, 118, 358, 127 ], "score": 0.3, "content": "\\mathbf { B }", "type": "inline_equation" }, { "bbox": [ 358, 117, 505, 129 ], "score": 1.0, "content": "and product C, each can be further", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 128, 505, 140 ], "spans": [ { "bbox": [ 105, 128, 505, 140 ], "score": 1.0, "content": "classified into three foreground domains—Normal, Scratches and Spots. Figure 7 shows", "type": "text" } ], "index": 3 }, { "bbox": [ 104, 138, 505, 153 ], "spans": [ { "bbox": [ 104, 138, 505, 153 ], "score": 1.0, "content": "example images of the SDI dataset. To be noticed that the dataset is highly imbalanced not only", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 150, 504, 161 ], "spans": [ { "bbox": [ 106, 150, 504, 161 ], "score": 1.0, "content": "between normal and defective samples but also between different products as shown in Table 5.", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 160, 506, 174 ], "spans": [ { "bbox": [ 105, 160, 506, 174 ], "score": 1.0, "content": "This sets a more challenging task when training deep neural networks like GANs and downstream", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 171, 151, 184 ], "spans": [ { "bbox": [ 105, 171, 151, 184 ], "score": 1.0, "content": "classifiers.", "type": "text" } ], "index": 7 } ], "index": 4, "bbox_fs": [ 104, 106, 506, 184 ] }, { "type": "text", "bbox": [ 107, 188, 505, 288 ], "lines": [ { "bbox": [ 105, 187, 505, 201 ], "spans": [ { "bbox": [ 105, 187, 505, 201 ], "score": 1.0, "content": "For each foreground and background domains, we randomly select 50 images for a joint valida-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 199, 505, 213 ], "spans": [ { "bbox": [ 105, 199, 505, 213 ], "score": 1.0, "content": "tion/test set, which is then further split into separate sets in the ratio of 3:7, and use all remaining", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 210, 506, 225 ], "spans": [ { "bbox": [ 105, 210, 506, 225 ], "score": 1.0, "content": "images as training sets for GAN and classifier training. We present the distribution of the training", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 222, 504, 235 ], "spans": [ { "bbox": [ 105, 222, 504, 235 ], "score": 1.0, "content": "set when training DT-GAN in Table 6. Note that the normal samples used in GAN training are only", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 232, 506, 245 ], "spans": [ { "bbox": [ 105, 232, 506, 245 ], "score": 1.0, "content": "a subset of all available samples in Normal and we keep the rest of them for generating defective", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 243, 505, 256 ], "spans": [ { "bbox": [ 105, 243, 505, 256 ], "score": 1.0, "content": "samples at test time. For classifier training, we show the statistics in Table 7, where the number", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 254, 506, 267 ], "spans": [ { "bbox": [ 105, 254, 506, 267 ], "score": 1.0, "content": "of normal samples involved in classifier training increase incrementally. The validation set is used", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 265, 505, 277 ], "spans": [ { "bbox": [ 105, 265, 505, 277 ], "score": 1.0, "content": "to select the best model during classifier training while the test set is left untouched until the final", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 277, 409, 289 ], "spans": [ { "bbox": [ 105, 277, 409, 289 ], "score": 1.0, "content": "evaluation. Both of the validation and test set are inaccessible by DT-GAN.", "type": "text" } ], "index": 16 } ], "index": 12, "bbox_fs": [ 105, 187, 506, 289 ] }, { "type": "table", "bbox": [ 118, 325, 297, 393 ], "blocks": [ { "type": "table_caption", "bbox": [ 120, 314, 295, 324 ], "group_id": 0, "lines": [ { "bbox": [ 118, 311, 297, 326 ], "spans": [ { "bbox": [ 118, 311, 297, 326 ], "score": 1.0, "content": "Table 5: Distribution of the full SDI dataset.", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "table_body", "bbox": [ 118, 325, 297, 393 ], "group_id": 0, "lines": [ { "bbox": [ 118, 325, 297, 393 ], "spans": [ { "bbox": [ 118, 325, 297, 393 ], "score": 0.979, "html": "
Overview
ABC
Normal625062506250
Scratches340167121
Spots108670258
", "type": "table", "image_path": "c5ad4cce383f42b999dbec5baa09abd0eec003f8449b17cca61eb4e7218407fe.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 118, 325, 297, 338.6 ], "spans": [], "index": 18 }, { "bbox": [ 118, 338.6, 297, 352.20000000000005 ], "spans": [], "index": 19 }, { "bbox": [ 118, 352.20000000000005, 297, 365.80000000000007 ], "spans": [], "index": 20 }, { "bbox": [ 118, 365.80000000000007, 297, 379.4000000000001 ], "spans": [], "index": 21 }, { "bbox": [ 118, 379.4000000000001, 297, 393.0000000000001 ], "spans": [], "index": 22 } ] } ], "index": 18.5 }, { "type": "table", "bbox": [ 315, 330, 495, 399 ], "blocks": [ { "type": "table_caption", "bbox": [ 314, 307, 495, 329 ], "group_id": 2, "lines": [ { "bbox": [ 315, 307, 496, 319 ], "spans": [ { "bbox": [ 315, 307, 496, 319 ], "score": 1.0, "content": "Table 6: The training set for DT-GAN and", "type": "text" } ], "index": 23 }, { "bbox": [ 314, 318, 469, 331 ], "spans": [ { "bbox": [ 314, 318, 469, 331 ], "score": 1.0, "content": "the baseline image synthesis methods.", "type": "text" } ], "index": 24 } ], "index": 23.5 }, { "type": "table_body", "bbox": [ 315, 330, 495, 399 ], "group_id": 2, "lines": [ { "bbox": [ 315, 330, 495, 399 ], "spans": [ { "bbox": [ 315, 330, 495, 399 ], "score": 0.977, "html": "
Overview
ABC
Normal700700700
Scratches29011771
Spots58620208
", "type": "table", "image_path": "d3b1a834deec1f06f6598262b6c7958b3d090165d454d73f8326527e6462444b.jpg" } ] } ], "index": 27, "virtual_lines": [ { "bbox": [ 315, 330, 495, 343.8 ], "spans": [], "index": 25 }, { "bbox": [ 315, 343.8, 495, 357.6 ], "spans": [], "index": 26 }, { "bbox": [ 315, 357.6, 495, 371.40000000000003 ], "spans": [], "index": 27 }, { "bbox": [ 315, 371.40000000000003, 495, 385.20000000000005 ], "spans": [], "index": 28 }, { "bbox": [ 315, 385.20000000000005, 495, 399.00000000000006 ], "spans": [], "index": 29 } ] } ], "index": 25.25 }, { "type": "table", "bbox": [ 106, 447, 510, 517 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 425, 505, 446 ], "group_id": 1, "lines": [ { "bbox": [ 105, 423, 505, 438 ], "spans": [ { "bbox": [ 105, 423, 450, 438 ], "score": 1.0, "content": "Table 7: The training, validation and test set for classifier training, where", "type": "text" }, { "bbox": [ 451, 425, 461, 435 ], "score": 0.71, "content": "N", "type": "inline_equation" }, { "bbox": [ 462, 423, 505, 438 ], "score": 1.0, "content": "increases", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 435, 284, 447 ], "spans": [ { "bbox": [ 106, 435, 284, 447 ], "score": 1.0, "content": "incrementally—1500, 2200, 4000 and 6200.", "type": "text" } ], "index": 31 } ], "index": 30.5 }, { "type": "table_body", "bbox": [ 106, 447, 510, 517 ], "group_id": 1, "lines": [ { "bbox": [ 106, 447, 510, 517 ], "spans": [ { "bbox": [ 106, 447, 510, 517 ], "score": 0.978, "html": "
TrainValidationTest
ABCABCABC
NormalNNN121815383235
Scratches29011771141615363435
Spots58620208141615363435
", "type": "table", "image_path": "ecde65a45fe9d34b0e307da325180a667544009c522094149b40f88f28ea42b7.jpg" } ] } ], "index": 33, "virtual_lines": [ { "bbox": [ 106, 447, 510, 470.3333333333333 ], "spans": [], "index": 32 }, { "bbox": [ 106, 470.3333333333333, 510, 493.66666666666663 ], "spans": [], "index": 33 }, { "bbox": [ 106, 493.66666666666663, 510, 517.0 ], "spans": [], "index": 34 } ] } ], "index": 31.75 }, { "type": "title", "bbox": [ 108, 539, 228, 552 ], "lines": [ { "bbox": [ 105, 538, 230, 554 ], "spans": [ { "bbox": [ 105, 538, 230, 554 ], "score": 1.0, "content": "B TRAINING DETAILS", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 106, 564, 505, 630 ], "lines": [ { "bbox": [ 105, 563, 506, 576 ], "spans": [ { "bbox": [ 105, 563, 506, 576 ], "score": 1.0, "content": "DT-GAN. We follow the training scheme as described in StarGAN v2 with minor modifications. To", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 575, 505, 587 ], "spans": [ { "bbox": [ 105, 575, 505, 587 ], "score": 1.0, "content": "fit the model on a single Nvidia GTX TITAN X, the batch size is reduced to four while the model is", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 586, 506, 599 ], "spans": [ { "bbox": [ 105, 586, 506, 599 ], "score": 1.0, "content": "still trained for 100,000 iterations. The training time is about three and a half days on the dedicated", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 597, 504, 609 ], "spans": [ { "bbox": [ 106, 597, 504, 609 ], "score": 1.0, "content": "GPU with the modified network architecture2 and loss functions mentioned in Section 3 in PyTorch", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 607, 506, 621 ], "spans": [ { "bbox": [ 105, 607, 224, 621 ], "score": 1.0, "content": "(Paszke et al., 2017). We set", "type": "text" }, { "bbox": [ 224, 608, 255, 620 ], "score": 0.89, "content": "\\lambda _ { \\mathrm { s t y } } = 1", "type": "inline_equation" }, { "bbox": [ 256, 607, 259, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 260, 608, 289, 619 ], "score": 0.86, "content": "\\lambda _ { \\mathrm { d s } } = 1", "type": "inline_equation" }, { "bbox": [ 290, 607, 293, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 294, 608, 327, 620 ], "score": 0.88, "content": "\\lambda _ { \\mathrm { c y c } } = 1", "type": "inline_equation" }, { "bbox": [ 327, 607, 331, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 331, 608, 377, 620 ], "score": 0.86, "content": "\\lambda _ { \\mathrm { c o n . c y c } } = 1", "type": "inline_equation" }, { "bbox": [ 377, 607, 380, 621 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 381, 608, 412, 619 ], "score": 0.89, "content": "\\lambda _ { \\mathrm { c l s } } = 1", "type": "inline_equation" }, { "bbox": [ 413, 607, 431, 621 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 431, 608, 474, 619 ], "score": 0.92, "content": "\\lambda _ { \\mathrm { B G . c l s } } = 1", "type": "inline_equation" }, { "bbox": [ 475, 607, 506, 621 ], "score": 1.0, "content": "for the", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 618, 402, 631 ], "spans": [ { "bbox": [ 105, 618, 402, 631 ], "score": 1.0, "content": "SDI dataset. All other design choices remain the same as in StarGAN v2.", "type": "text" } ], "index": 41 } ], "index": 38.5, "bbox_fs": [ 105, 563, 506, 631 ] }, { "type": "text", "bbox": [ 106, 635, 505, 713 ], "lines": [ { "bbox": [ 106, 636, 505, 648 ], "spans": [ { "bbox": [ 106, 636, 505, 648 ], "score": 1.0, "content": "Classifiers. We train all the classifiers that use ResNet-50 as backbone for 100 epochs with the", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 646, 506, 659 ], "spans": [ { "bbox": [ 105, 646, 506, 659 ], "score": 1.0, "content": "SGD optimizer (Ruder, 2016) and batch size 256. The initial learning rate is 0.001, momentum is", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 658, 505, 670 ], "spans": [ { "bbox": [ 105, 658, 505, 670 ], "score": 1.0, "content": "0.9 and weight decay is 1e-4. A learning rate scheduler is set to reduce the learning rate by factor", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 668, 505, 681 ], "spans": [ { "bbox": [ 105, 668, 505, 681 ], "score": 1.0, "content": "of 0.1 when the validation loss stops decreasing for 5 epochs. The same setting also applies to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 679, 505, 692 ], "spans": [ { "bbox": [ 105, 679, 505, 692 ], "score": 1.0, "content": "EfficientNet-b4, except the batch size is reduced to 128. Although DT-GAN can synthesize realistic", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 691, 506, 704 ], "spans": [ { "bbox": [ 105, 691, 506, 704 ], "score": 1.0, "content": "defective samples, we notice that there still exists a domain gap between the generated samples", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 702, 505, 716 ], "spans": [ { "bbox": [ 105, 702, 505, 716 ], "score": 1.0, "content": "and the real samples. To explore the full potential of the generated samples, we attach an auxiliary", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 616, 505, 627 ], "spans": [ { "bbox": [ 105, 616, 505, 627 ], "score": 1.0, "content": "source classifier to distinguish between synthetic and real samples. Then, this classifier is connected", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 105, 627, 505, 639 ], "spans": [ { "bbox": [ 105, 627, 505, 639 ], "score": 1.0, "content": "to the backbone (e.g. ResNet-50) through a Gradient Reversal Layer. With the help of the Gradient", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 105, 638, 505, 650 ], "spans": [ { "bbox": [ 105, 638, 505, 650 ], "score": 1.0, "content": "Reversal Layer, the backbone is forced to extract the shared features between synthetic and real", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 105, 649, 377, 661 ], "spans": [ { "bbox": [ 105, 649, 377, 661 ], "score": 1.0, "content": "samples, which ensures all training samples are effectively learned.", "type": "text", "cross_page": true } ], "index": 7 } ], "index": 45, "bbox_fs": [ 105, 636, 506, 716 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 66, 506, 526 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 66, 506, 526 ], "group_id": 0, "lines": [ { "bbox": [ 106, 66, 506, 526 ], "spans": [ { "bbox": [ 106, 66, 506, 526 ], "score": 0.853, "type": "image", "image_path": "2b5df2b8d264369f1ee04ca1e3d7919ee1d3d69e7370aa9deec145bb6941e15e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 66, 506, 219.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 106, 219.33333333333334, 506, 372.6666666666667 ], "spans": [], "index": 1 }, { "bbox": [ 106, 372.6666666666667, 506, 526.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 227, 538, 384, 550 ], "group_id": 0, "lines": [ { "bbox": [ 226, 537, 384, 552 ], "spans": [ { "bbox": [ 226, 537, 384, 552 ], "score": 1.0, "content": "Figure 7: Overview of the SDI dataset.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "text", "bbox": [ 107, 615, 505, 660 ], "lines": [ { "bbox": [ 105, 616, 505, 627 ], "spans": [ { "bbox": [ 105, 616, 505, 627 ], "score": 1.0, "content": "source classifier to distinguish between synthetic and real samples. Then, this classifier is connected", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 627, 505, 639 ], "spans": [ { "bbox": [ 105, 627, 505, 639 ], "score": 1.0, "content": "to the backbone (e.g. ResNet-50) through a Gradient Reversal Layer. With the help of the Gradient", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 638, 505, 650 ], "spans": [ { "bbox": [ 105, 638, 505, 650 ], "score": 1.0, "content": "Reversal Layer, the backbone is forced to extract the shared features between synthetic and real", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 649, 377, 661 ], "spans": [ { "bbox": [ 105, 649, 377, 661 ], "score": 1.0, "content": "samples, which ensures all training samples are effectively learned.", "type": "text" } ], "index": 7 } ], "index": 5.5 }, { "type": "text", "bbox": [ 106, 666, 505, 732 ], "lines": [ { "bbox": [ 106, 665, 505, 678 ], "spans": [ { "bbox": [ 106, 665, 505, 678 ], "score": 1.0, "content": "We design a two-layer perceptron that connects to the average pooling layer in ResNet-50 as shown", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 677, 505, 690 ], "spans": [ { "bbox": [ 105, 677, 505, 690 ], "score": 1.0, "content": "in Figure 8. Note that the usual fully connected layer after the average pooling in ResNet-50 remains", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 687, 505, 700 ], "spans": [ { "bbox": [ 105, 687, 505, 700 ], "score": 1.0, "content": "the same and is not affected by the extra branch we added. Inspired by Chen et al. (2018), a three-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "layer perceptron is used for EfficientNet-b4 instead as shown in Figure 9. Its layers are initialized", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "with a random normal distribution, where the standard deviation is set to 0.01 for the first two layers", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 721, 371, 732 ], "spans": [ { "bbox": [ 106, 721, 371, 732 ], "score": 1.0, "content": "and 0.05 for the output layer. The biases for all layers are set to 0.", "type": "text" } ], "index": 13 } ], "index": 10.5 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "14", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 66, 506, 526 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 66, 506, 526 ], "group_id": 0, "lines": [ { "bbox": [ 106, 66, 506, 526 ], "spans": [ { "bbox": [ 106, 66, 506, 526 ], "score": 0.853, "type": "image", "image_path": "2b5df2b8d264369f1ee04ca1e3d7919ee1d3d69e7370aa9deec145bb6941e15e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 66, 506, 219.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 106, 219.33333333333334, 506, 372.6666666666667 ], "spans": [], "index": 1 }, { "bbox": [ 106, 372.6666666666667, 506, 526.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 227, 538, 384, 550 ], "group_id": 0, "lines": [ { "bbox": [ 226, 537, 384, 552 ], "spans": [ { "bbox": [ 226, 537, 384, 552 ], "score": 1.0, "content": "Figure 7: Overview of the SDI dataset.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 }, { "type": "text", "bbox": [ 107, 615, 505, 660 ], "lines": [], "index": 5.5, "bbox_fs": [ 105, 616, 505, 661 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 666, 505, 732 ], "lines": [ { "bbox": [ 106, 665, 505, 678 ], "spans": [ { "bbox": [ 106, 665, 505, 678 ], "score": 1.0, "content": "We design a two-layer perceptron that connects to the average pooling layer in ResNet-50 as shown", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 677, 505, 690 ], "spans": [ { "bbox": [ 105, 677, 505, 690 ], "score": 1.0, "content": "in Figure 8. Note that the usual fully connected layer after the average pooling in ResNet-50 remains", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 687, 505, 700 ], "spans": [ { "bbox": [ 105, 687, 505, 700 ], "score": 1.0, "content": "the same and is not affected by the extra branch we added. Inspired by Chen et al. (2018), a three-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "layer perceptron is used for EfficientNet-b4 instead as shown in Figure 9. Its layers are initialized", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "with a random normal distribution, where the standard deviation is set to 0.01 for the first two layers", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 721, 371, 732 ], "spans": [ { "bbox": [ 106, 721, 371, 732 ], "score": 1.0, "content": "and 0.05 for the output layer. The biases for all layers are set to 0.", "type": "text" } ], "index": 13 } ], "index": 10.5, "bbox_fs": [ 105, 665, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 130, 81, 282, 133 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 81, 282, 133 ], "group_id": 0, "lines": [ { "bbox": [ 130, 81, 282, 133 ], "spans": [ { "bbox": [ 130, 81, 282, 133 ], "score": 0.957, "type": "image", "image_path": "5876fcb93af3f0aad7fc6536e733bff52799290b671bdeef4fd81bf21ab172fb.jpg" } ] } ], "index": 1.5, "virtual_lines": [ { "bbox": [ 130, 81, 282, 107.0 ], "spans": [], "index": 0 }, { "bbox": [ 130, 107.0, 282, 133.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 142, 146, 271, 158 ], "group_id": 0, "lines": [ { "bbox": [ 141, 145, 271, 159 ], "spans": [ { "bbox": [ 141, 145, 271, 159 ], "score": 1.0, "content": "Figure 8: ResNet-50 with GRL.", "type": "text" } ], "index": 6 } ], "index": 6 } ], "index": 3.75 }, { "type": "image", "bbox": [ 306, 81, 506, 133 ], "blocks": [ { "type": "image_body", "bbox": [ 306, 81, 506, 133 ], "group_id": 1, "lines": [ { "bbox": [ 306, 81, 506, 133 ], "spans": [ { "bbox": [ 306, 81, 506, 133 ], "score": 0.95, "type": "image", "image_path": "79d423c36937100d387ec0e803940f1adbc96236a1a0b915320ec3741d5f7095.jpg" } ] } ], "index": 3.0, "virtual_lines": [ { "bbox": [ 306, 81, 506, 94.0 ], "spans": [], "index": 1 }, { "bbox": [ 306, 94.0, 506, 107.0 ], "spans": [], "index": 2 }, { "bbox": [ 306, 107.0, 506, 120.0 ], "spans": [], "index": 4 }, { "bbox": [ 306, 120.0, 506, 133.0 ], "spans": [], "index": 5 } ] }, { "type": "image_caption", "bbox": [ 333, 146, 480, 158 ], "group_id": 1, "lines": [ { "bbox": [ 332, 145, 480, 159 ], "spans": [ { "bbox": [ 332, 145, 480, 159 ], "score": 1.0, "content": "Figure 9: EfficientNet-b4 with GRL.", "type": "text" } ], "index": 7 } ], "index": 7 } ], "index": 5.0 }, { "type": "title", "bbox": [ 108, 175, 231, 187 ], "lines": [ { "bbox": [ 106, 173, 232, 189 ], "spans": [ { "bbox": [ 106, 173, 232, 189 ], "score": 1.0, "content": "C EVALUATION SETUP", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 106, 199, 505, 298 ], "lines": [ { "bbox": [ 106, 199, 505, 212 ], "spans": [ { "bbox": [ 106, 199, 505, 212 ], "score": 1.0, "content": "Generated samples from DT-GAN. DT-GAN requires images as input for generating synthetic", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 209, 505, 223 ], "spans": [ { "bbox": [ 105, 209, 505, 223 ], "score": 1.0, "content": "data. At test time, we translated each Normal image in the SDI dataset into four defective images:", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 221, 504, 234 ], "spans": [ { "bbox": [ 105, 221, 504, 234 ], "score": 1.0, "content": "two with Scratches and two with Spots. The translations were performed by two subnetworks:", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 232, 505, 245 ], "spans": [ { "bbox": [ 105, 232, 207, 245 ], "score": 1.0, "content": "by the mapping network", "type": "text" }, { "bbox": [ 208, 233, 219, 243 ], "score": 0.69, "content": "M", "type": "inline_equation" }, { "bbox": [ 220, 232, 505, 245 ], "score": 1.0, "content": "using random noise (‘latent-guided’) and by the style-content encoder", "type": "text" } ], "index": 12 }, { "bbox": [ 107, 243, 505, 256 ], "spans": [ { "bbox": [ 107, 243, 116, 253 ], "score": 0.78, "content": "E", "type": "inline_equation" }, { "bbox": [ 116, 243, 505, 256 ], "score": 1.0, "content": "using a reference image (‘reference-guided’). We first randomly sampled one latent code for", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 254, 505, 267 ], "spans": [ { "bbox": [ 106, 254, 505, 267 ], "score": 1.0, "content": "each defective foreground domain. Similarly, we also randomly sampled one reference image from", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 265, 505, 277 ], "spans": [ { "bbox": [ 106, 265, 505, 277 ], "score": 1.0, "content": "the training set for each defective foreground domain. The corresponding style codes and defect", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 275, 506, 290 ], "spans": [ { "bbox": [ 104, 275, 506, 290 ], "score": 1.0, "content": "contents were then produced by the two subnetworks respectively and fed to the generator for target", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 288, 180, 300 ], "spans": [ { "bbox": [ 106, 288, 180, 300 ], "score": 1.0, "content": "image generation.", "type": "text" } ], "index": 17 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 303, 505, 403 ], "lines": [ { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 505, 316 ], "score": 1.0, "content": "We conducted classification experiments separately on images generated from the two subnetworks", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 315, 505, 327 ], "spans": [ { "bbox": [ 106, 316, 235, 327 ], "score": 1.0, "content": "and a mixture set of both (i.e.", "type": "text" }, { "bbox": [ 235, 315, 255, 326 ], "score": 0.85, "content": "50 \\%", "type": "inline_equation" }, { "bbox": [ 256, 316, 505, 327 ], "score": 1.0, "content": "from each subnetwork). Experiments show consistent gains", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 325, 506, 339 ], "spans": [ { "bbox": [ 105, 325, 506, 339 ], "score": 1.0, "content": "of using synthetic images generated from DT-GAN (Table 8). We observe that the latent-guided", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 337, 506, 349 ], "spans": [ { "bbox": [ 105, 337, 506, 349 ], "score": 1.0, "content": "synthetic images in general perform better than the reference-guided one, while the mixture set", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 348, 506, 361 ], "spans": [ { "bbox": [ 105, 348, 506, 361 ], "score": 1.0, "content": "provides more stable results with regard to the standard deviation. Presumably the mixture set", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 358, 505, 371 ], "spans": [ { "bbox": [ 105, 358, 505, 371 ], "score": 1.0, "content": "benefits from the combination of samples from reference-guided synthesis, which are well aligned", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 370, 505, 382 ], "spans": [ { "bbox": [ 106, 370, 505, 382 ], "score": 1.0, "content": "with the original defect distribution, and the samples from latent-guided synthesis, i.e. from random", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 381, 505, 393 ], "spans": [ { "bbox": [ 106, 381, 505, 393 ], "score": 1.0, "content": "noise, which adds novel but plausible defects to the dataset. In the main text, we report the results", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 392, 460, 404 ], "spans": [ { "bbox": [ 105, 392, 460, 404 ], "score": 1.0, "content": "of the mixture set for all experiments, including the quantitative evaluation of DT-GAN.", "type": "text" } ], "index": 26 } ], "index": 22 }, { "type": "table", "bbox": [ 175, 442, 434, 522 ], "blocks": [ { "type": "table_caption", "bbox": [ 104, 421, 502, 442 ], "group_id": 0, "lines": [ { "bbox": [ 105, 421, 505, 434 ], "spans": [ { "bbox": [ 105, 421, 505, 434 ], "score": 1.0, "content": "Table 8: Classification results with regard to the synthetic images generated from the two subnet-", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 432, 214, 444 ], "spans": [ { "bbox": [ 106, 432, 214, 444 ], "score": 1.0, "content": "works and the mixture set.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "table_body", "bbox": [ 175, 442, 434, 522 ], "group_id": 0, "lines": [ { "bbox": [ 175, 442, 434, 522 ], "spans": [ { "bbox": [ 175, 442, 434, 522 ], "score": 0.981, "html": "
Dataset SizeAll
Trad-AugLatentReferenceMix
450012.75±0.6110.72±0.9611.48±0.8811.04±0.76
660013.07±1.5710.34±1.8611.55±1.6410.60±0.48
1200012.05±0.819.90±1.2610.40±0.999.90±0.69
1860012.37±0.3211.04±1.2612.12±0.7510.21±0.96
", "type": "table", "image_path": "c1d4e75efc50b8836e9f56d1642ae073df8bae94b456b2fba3105e400e5d6572.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 175, 442, 434, 468.6666666666667 ], "spans": [], "index": 29 }, { "bbox": [ 175, 468.6666666666667, 434, 495.33333333333337 ], "spans": [], "index": 30 }, { "bbox": [ 175, 495.33333333333337, 434, 522.0 ], "spans": [], "index": 31 } ] } ], "index": 28.75 }, { "type": "text", "bbox": [ 106, 537, 505, 680 ], "lines": [ { "bbox": [ 106, 537, 506, 550 ], "spans": [ { "bbox": [ 106, 537, 506, 550 ], "score": 1.0, "content": "Frechet inception distance (FID) and Kernel inception distance (KID). ´ We used the feature", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 549, 505, 561 ], "spans": [ { "bbox": [ 106, 549, 505, 561 ], "score": 1.0, "content": "vectors from the last average pooling layer of the ImageNet pretrained Inception-V3 to calculate", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 559, 505, 573 ], "spans": [ { "bbox": [ 105, 559, 505, 573 ], "score": 1.0, "content": "both scores. For each test image from the Normal domain, we translated it into a synthetic defective", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 570, 505, 583 ], "spans": [ { "bbox": [ 105, 570, 505, 583 ], "score": 1.0, "content": "image of each defect domain. The style codes and contents for the translation were acquired in two", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 581, 506, 594 ], "spans": [ { "bbox": [ 106, 581, 506, 594 ], "score": 1.0, "content": "ways: by randomly sampling from the standard normal distribution and by randomly sampling a", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 592, 506, 605 ], "spans": [ { "bbox": [ 105, 592, 506, 605 ], "score": 1.0, "content": "reference image from the train set of a defect domain. To calculate the FID and KID score, we", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 603, 506, 617 ], "spans": [ { "bbox": [ 104, 603, 506, 617 ], "score": 1.0, "content": "generated 4000 defective samples per product per defect domain for each way of guidance, and", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 614, 506, 627 ], "spans": [ { "bbox": [ 106, 614, 506, 627 ], "score": 1.0, "content": "formed the mixture set by randomly sampling 2000 images per product per defect domain from", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 625, 506, 638 ], "spans": [ { "bbox": [ 106, 625, 506, 638 ], "score": 1.0, "content": "each way. The reported FID and KID scores were then computed between the defective images in", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 635, 506, 649 ], "spans": [ { "bbox": [ 105, 635, 506, 649 ], "score": 1.0, "content": "the training set and the mixture set of synthetic defective images. The same procedure was applied", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 646, 505, 659 ], "spans": [ { "bbox": [ 106, 646, 505, 659 ], "score": 1.0, "content": "when computing scores on single product subsets of the SDI dataset. For example, for product A,", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 658, 505, 671 ], "spans": [ { "bbox": [ 106, 658, 505, 671 ], "score": 1.0, "content": "we calculated the scores between the defective image of product A in the training set and the mixture", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 669, 295, 682 ], "spans": [ { "bbox": [ 106, 669, 295, 682 ], "score": 1.0, "content": "set of synthetic defective images of product A.", "type": "text" } ], "index": 44 } ], "index": 38 }, { "type": "title", "bbox": [ 108, 696, 267, 709 ], "lines": [ { "bbox": [ 106, 695, 268, 710 ], "spans": [ { "bbox": [ 106, 695, 268, 710 ], "score": 1.0, "content": "D NETWORK ARCHITECTURE", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 106, 720, 439, 732 ], "lines": [ { "bbox": [ 105, 719, 441, 733 ], "spans": [ { "bbox": [ 105, 719, 441, 733 ], "score": 1.0, "content": "In this section, we provide the architectural details of all four modules in DT-GAN.", "type": "text" } ], "index": 46 } ], "index": 46 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 13 } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 130, 81, 282, 133 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 81, 282, 133 ], "group_id": 0, "lines": [ { "bbox": [ 130, 81, 282, 133 ], "spans": [ { "bbox": [ 130, 81, 282, 133 ], "score": 0.957, "type": "image", "image_path": "5876fcb93af3f0aad7fc6536e733bff52799290b671bdeef4fd81bf21ab172fb.jpg" } ] } ], "index": 1.5, "virtual_lines": [ { "bbox": [ 130, 81, 282, 107.0 ], "spans": [], "index": 0 }, { "bbox": [ 130, 107.0, 282, 133.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 142, 146, 271, 158 ], "group_id": 0, "lines": [ { "bbox": [ 141, 145, 271, 159 ], "spans": [ { "bbox": [ 141, 145, 271, 159 ], "score": 1.0, "content": "Figure 8: ResNet-50 with GRL.", "type": "text" } ], "index": 6 } ], "index": 6 } ], "index": 3.75 }, { "type": "image", "bbox": [ 306, 81, 506, 133 ], "blocks": [ { "type": "image_body", "bbox": [ 306, 81, 506, 133 ], "group_id": 1, "lines": [ { "bbox": [ 306, 81, 506, 133 ], "spans": [ { "bbox": [ 306, 81, 506, 133 ], "score": 0.95, "type": "image", "image_path": "79d423c36937100d387ec0e803940f1adbc96236a1a0b915320ec3741d5f7095.jpg" } ] } ], "index": 3.0, "virtual_lines": [ { "bbox": [ 306, 81, 506, 94.0 ], "spans": [], "index": 1 }, { "bbox": [ 306, 94.0, 506, 107.0 ], "spans": [], "index": 2 }, { "bbox": [ 306, 107.0, 506, 120.0 ], "spans": [], "index": 4 }, { "bbox": [ 306, 120.0, 506, 133.0 ], "spans": [], "index": 5 } ] }, { "type": "image_caption", "bbox": [ 333, 146, 480, 158 ], "group_id": 1, "lines": [ { "bbox": [ 332, 145, 480, 159 ], "spans": [ { "bbox": [ 332, 145, 480, 159 ], "score": 1.0, "content": "Figure 9: EfficientNet-b4 with GRL.", "type": "text" } ], "index": 7 } ], "index": 7 } ], "index": 5.0 }, { "type": "title", "bbox": [ 108, 175, 231, 187 ], "lines": [ { "bbox": [ 106, 173, 232, 189 ], "spans": [ { "bbox": [ 106, 173, 232, 189 ], "score": 1.0, "content": "C EVALUATION SETUP", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 106, 199, 505, 298 ], "lines": [ { "bbox": [ 106, 199, 505, 212 ], "spans": [ { "bbox": [ 106, 199, 505, 212 ], "score": 1.0, "content": "Generated samples from DT-GAN. DT-GAN requires images as input for generating synthetic", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 209, 505, 223 ], "spans": [ { "bbox": [ 105, 209, 505, 223 ], "score": 1.0, "content": "data. At test time, we translated each Normal image in the SDI dataset into four defective images:", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 221, 504, 234 ], "spans": [ { "bbox": [ 105, 221, 504, 234 ], "score": 1.0, "content": "two with Scratches and two with Spots. The translations were performed by two subnetworks:", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 232, 505, 245 ], "spans": [ { "bbox": [ 105, 232, 207, 245 ], "score": 1.0, "content": "by the mapping network", "type": "text" }, { "bbox": [ 208, 233, 219, 243 ], "score": 0.69, "content": "M", "type": "inline_equation" }, { "bbox": [ 220, 232, 505, 245 ], "score": 1.0, "content": "using random noise (‘latent-guided’) and by the style-content encoder", "type": "text" } ], "index": 12 }, { "bbox": [ 107, 243, 505, 256 ], "spans": [ { "bbox": [ 107, 243, 116, 253 ], "score": 0.78, "content": "E", "type": "inline_equation" }, { "bbox": [ 116, 243, 505, 256 ], "score": 1.0, "content": "using a reference image (‘reference-guided’). We first randomly sampled one latent code for", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 254, 505, 267 ], "spans": [ { "bbox": [ 106, 254, 505, 267 ], "score": 1.0, "content": "each defective foreground domain. Similarly, we also randomly sampled one reference image from", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 265, 505, 277 ], "spans": [ { "bbox": [ 106, 265, 505, 277 ], "score": 1.0, "content": "the training set for each defective foreground domain. The corresponding style codes and defect", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 275, 506, 290 ], "spans": [ { "bbox": [ 104, 275, 506, 290 ], "score": 1.0, "content": "contents were then produced by the two subnetworks respectively and fed to the generator for target", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 288, 180, 300 ], "spans": [ { "bbox": [ 106, 288, 180, 300 ], "score": 1.0, "content": "image generation.", "type": "text" } ], "index": 17 } ], "index": 13, "bbox_fs": [ 104, 199, 506, 300 ] }, { "type": "text", "bbox": [ 107, 303, 505, 403 ], "lines": [ { "bbox": [ 106, 303, 505, 316 ], "spans": [ { "bbox": [ 106, 303, 505, 316 ], "score": 1.0, "content": "We conducted classification experiments separately on images generated from the two subnetworks", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 315, 505, 327 ], "spans": [ { "bbox": [ 106, 316, 235, 327 ], "score": 1.0, "content": "and a mixture set of both (i.e.", "type": "text" }, { "bbox": [ 235, 315, 255, 326 ], "score": 0.85, "content": "50 \\%", "type": "inline_equation" }, { "bbox": [ 256, 316, 505, 327 ], "score": 1.0, "content": "from each subnetwork). Experiments show consistent gains", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 325, 506, 339 ], "spans": [ { "bbox": [ 105, 325, 506, 339 ], "score": 1.0, "content": "of using synthetic images generated from DT-GAN (Table 8). We observe that the latent-guided", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 337, 506, 349 ], "spans": [ { "bbox": [ 105, 337, 506, 349 ], "score": 1.0, "content": "synthetic images in general perform better than the reference-guided one, while the mixture set", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 348, 506, 361 ], "spans": [ { "bbox": [ 105, 348, 506, 361 ], "score": 1.0, "content": "provides more stable results with regard to the standard deviation. Presumably the mixture set", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 358, 505, 371 ], "spans": [ { "bbox": [ 105, 358, 505, 371 ], "score": 1.0, "content": "benefits from the combination of samples from reference-guided synthesis, which are well aligned", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 370, 505, 382 ], "spans": [ { "bbox": [ 106, 370, 505, 382 ], "score": 1.0, "content": "with the original defect distribution, and the samples from latent-guided synthesis, i.e. from random", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 381, 505, 393 ], "spans": [ { "bbox": [ 106, 381, 505, 393 ], "score": 1.0, "content": "noise, which adds novel but plausible defects to the dataset. In the main text, we report the results", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 392, 460, 404 ], "spans": [ { "bbox": [ 105, 392, 460, 404 ], "score": 1.0, "content": "of the mixture set for all experiments, including the quantitative evaluation of DT-GAN.", "type": "text" } ], "index": 26 } ], "index": 22, "bbox_fs": [ 105, 303, 506, 404 ] }, { "type": "table", "bbox": [ 175, 442, 434, 522 ], "blocks": [ { "type": "table_caption", "bbox": [ 104, 421, 502, 442 ], "group_id": 0, "lines": [ { "bbox": [ 105, 421, 505, 434 ], "spans": [ { "bbox": [ 105, 421, 505, 434 ], "score": 1.0, "content": "Table 8: Classification results with regard to the synthetic images generated from the two subnet-", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 432, 214, 444 ], "spans": [ { "bbox": [ 106, 432, 214, 444 ], "score": 1.0, "content": "works and the mixture set.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "table_body", "bbox": [ 175, 442, 434, 522 ], "group_id": 0, "lines": [ { "bbox": [ 175, 442, 434, 522 ], "spans": [ { "bbox": [ 175, 442, 434, 522 ], "score": 0.981, "html": "
Dataset SizeAll
Trad-AugLatentReferenceMix
450012.75±0.6110.72±0.9611.48±0.8811.04±0.76
660013.07±1.5710.34±1.8611.55±1.6410.60±0.48
1200012.05±0.819.90±1.2610.40±0.999.90±0.69
1860012.37±0.3211.04±1.2612.12±0.7510.21±0.96
", "type": "table", "image_path": "c1d4e75efc50b8836e9f56d1642ae073df8bae94b456b2fba3105e400e5d6572.jpg" } ] } ], "index": 30, "virtual_lines": [ { "bbox": [ 175, 442, 434, 468.6666666666667 ], "spans": [], "index": 29 }, { "bbox": [ 175, 468.6666666666667, 434, 495.33333333333337 ], "spans": [], "index": 30 }, { "bbox": [ 175, 495.33333333333337, 434, 522.0 ], "spans": [], "index": 31 } ] } ], "index": 28.75 }, { "type": "text", "bbox": [ 106, 537, 505, 680 ], "lines": [ { "bbox": [ 106, 537, 506, 550 ], "spans": [ { "bbox": [ 106, 537, 506, 550 ], "score": 1.0, "content": "Frechet inception distance (FID) and Kernel inception distance (KID). ´ We used the feature", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 549, 505, 561 ], "spans": [ { "bbox": [ 106, 549, 505, 561 ], "score": 1.0, "content": "vectors from the last average pooling layer of the ImageNet pretrained Inception-V3 to calculate", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 559, 505, 573 ], "spans": [ { "bbox": [ 105, 559, 505, 573 ], "score": 1.0, "content": "both scores. For each test image from the Normal domain, we translated it into a synthetic defective", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 570, 505, 583 ], "spans": [ { "bbox": [ 105, 570, 505, 583 ], "score": 1.0, "content": "image of each defect domain. The style codes and contents for the translation were acquired in two", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 581, 506, 594 ], "spans": [ { "bbox": [ 106, 581, 506, 594 ], "score": 1.0, "content": "ways: by randomly sampling from the standard normal distribution and by randomly sampling a", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 592, 506, 605 ], "spans": [ { "bbox": [ 105, 592, 506, 605 ], "score": 1.0, "content": "reference image from the train set of a defect domain. To calculate the FID and KID score, we", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 603, 506, 617 ], "spans": [ { "bbox": [ 104, 603, 506, 617 ], "score": 1.0, "content": "generated 4000 defective samples per product per defect domain for each way of guidance, and", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 614, 506, 627 ], "spans": [ { "bbox": [ 106, 614, 506, 627 ], "score": 1.0, "content": "formed the mixture set by randomly sampling 2000 images per product per defect domain from", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 625, 506, 638 ], "spans": [ { "bbox": [ 106, 625, 506, 638 ], "score": 1.0, "content": "each way. The reported FID and KID scores were then computed between the defective images in", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 635, 506, 649 ], "spans": [ { "bbox": [ 105, 635, 506, 649 ], "score": 1.0, "content": "the training set and the mixture set of synthetic defective images. The same procedure was applied", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 646, 505, 659 ], "spans": [ { "bbox": [ 106, 646, 505, 659 ], "score": 1.0, "content": "when computing scores on single product subsets of the SDI dataset. For example, for product A,", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 658, 505, 671 ], "spans": [ { "bbox": [ 106, 658, 505, 671 ], "score": 1.0, "content": "we calculated the scores between the defective image of product A in the training set and the mixture", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 669, 295, 682 ], "spans": [ { "bbox": [ 106, 669, 295, 682 ], "score": 1.0, "content": "set of synthetic defective images of product A.", "type": "text" } ], "index": 44 } ], "index": 38, "bbox_fs": [ 104, 537, 506, 682 ] }, { "type": "title", "bbox": [ 108, 696, 267, 709 ], "lines": [ { "bbox": [ 106, 695, 268, 710 ], "spans": [ { "bbox": [ 106, 695, 268, 710 ], "score": 1.0, "content": "D NETWORK ARCHITECTURE", "type": "text" } ], "index": 45 } ], "index": 45 }, { "type": "text", "bbox": [ 106, 720, 439, 732 ], "lines": [ { "bbox": [ 105, 719, 441, 733 ], "spans": [ { "bbox": [ 105, 719, 441, 733 ], "score": 1.0, "content": "In this section, we provide the architectural details of all four modules in DT-GAN.", "type": "text" } ], "index": 46 } ], "index": 46, "bbox_fs": [ 105, 719, 441, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 108, 105, 501, 394 ], "blocks": [ { "type": "table_caption", "bbox": [ 241, 90, 369, 100 ], "group_id": 0, "lines": [ { "bbox": [ 240, 88, 370, 101 ], "spans": [ { "bbox": [ 240, 88, 370, 101 ], "score": 1.0, "content": "Table 9: Generator architecture.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 108, 105, 501, 394 ], "group_id": 0, "lines": [ { "bbox": [ 108, 105, 501, 394 ], "spans": [ { "bbox": [ 108, 105, 501, 394 ], "score": 0.979, "html": "
(a)Encoder
LayerResampleNormOutput Shape
Image x-128 × 128×3
Conv 1×1-128 ×128 ×128
ResBlkAvgPoolIN64× 64×256
ResBlkAvgPoolIN32 × 32 × 512
ResBlkAvgPoolIN16 × 16 × 512
ResBlkIN16 ×16× 512
ResBlkIN16 ×16 × 512
(b) Background Decoder(c) Foreground Decoder
LayerResampleNorm Output ShapeLayerResample NormOutput Shape
Input-16 × 16 × 448Input ResBlk16 × 16 × 64
ResBlkIN 16 ×16× 448 16 ×16× 512ResBlkAdaIN16 ×16× 64
ResBlkIN=AdaIN16 × 16 × 256
ResBlk=IN16 ×16 × 512ResBlkAdaIN16 ×16× 256
ResBlkUpsampleIN32 × 32×512ResBlk Upsample ResBlkAdaIN32 × 32 × 256
ResBlk ResBlkUpsample UpsampleIN IN64 × 64× 256 128 × 128× 448UpsampleAdaIN64×64×128
ResBlkUpsampleAdaIN128 ×128 × 64
(d) Fusion
LayerResampleNormOutput Shape
Input128 × 128 × (448 + 64)
Conv 1×1128 × 128× 3
", "type": "table", "image_path": "1c48ba4f6a9c5f47866cc9bb38143dabaa253479bdaae65b20f932c892e1f632.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 108, 105, 501, 201.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 108, 201.33333333333331, 501, 297.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 108, 297.66666666666663, 501, 393.99999999999994 ], "spans": [], "index": 3 } ] } ], "index": 1.0 }, { "type": "table", "bbox": [ 106, 439, 505, 637 ], "blocks": [ { "type": "table_caption", "bbox": [ 224, 421, 386, 432 ], "group_id": 1, "lines": [ { "bbox": [ 223, 420, 388, 433 ], "spans": [ { "bbox": [ 223, 420, 388, 433 ], "score": 1.0, "content": "Table 10: Mapping network architecture.", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "table_body", "bbox": [ 106, 439, 505, 637 ], "group_id": 1, "lines": [ { "bbox": [ 106, 439, 505, 637 ], "spans": [ { "bbox": [ 106, 439, 505, 637 ], "score": 0.966, "html": "
LayerActivationOutput Shape
Latent z=16
LinearReLU512
LinearReLU512
LinearReLU512
LinearReLU512
(b) Style Code(c) Content
LayerActivation Output ShapeLayerResample ActivationNoise Output Shape
Input=512Input=512
LinearReLU512Reshape--1×1×512
LinearReLU512ResBlkUpsampleINTrue2×2×512
LinearReLU512ResBlkUpsampleINTrue4×4×512
Linear164ResBlkUpsampleINTrue8×8×256
ResBlkUpsampleINTrue16 ×16×128
Conv 1×1=INTrue16 × 16× 64
", "type": "table", "image_path": "1098199c42bf1a6e4062025eaaa671f3e1525ccdf2528cdf76e0b9528b15881a.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 106, 439, 505, 505.0 ], "spans": [], "index": 5 }, { "bbox": [ 106, 505.0, 505, 571.0 ], "spans": [], "index": 6 }, { "bbox": [ 106, 571.0, 505, 637.0 ], "spans": [], "index": 7 } ] } ], "index": 5.0 }, { "type": "text", "bbox": [ 106, 665, 505, 732 ], "lines": [ { "bbox": [ 106, 666, 505, 678 ], "spans": [ { "bbox": [ 106, 666, 505, 678 ], "score": 1.0, "content": "Generator (Table 9). For the SDI dataset, the encoder part of the generator consists of three down-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "sampling blocks and two intermediate blocks (Table 9 (a)), all of them are pre-activation residual", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "units (He et al., 2016b). Then the encoded feature map is split channel-wise into background (Ta-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 698, 506, 711 ], "spans": [ { "bbox": [ 106, 698, 506, 711 ], "score": 1.0, "content": "ble 9 (b)) and foreground (Table 9 (c)). Both of them are then carried through separate decoders. We", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 710, 505, 721 ], "spans": [ { "bbox": [ 106, 710, 505, 721 ], "score": 1.0, "content": "use the instance normalization (IN) and the adaptive instance normalization (AdaIN) as indicated.", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 721, 506, 733 ], "spans": [ { "bbox": [ 106, 721, 506, 733 ], "score": 1.0, "content": "The style code is injected into all AdaIN layers to modulate the affine transformations. Note that", "type": "text" } ], "index": 13 } ], "index": 10.5 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "16", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 308, 38 ], "spans": [ { "bbox": [ 106, 25, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 108, 105, 501, 394 ], "blocks": [ { "type": "table_caption", "bbox": [ 241, 90, 369, 100 ], "group_id": 0, "lines": [ { "bbox": [ 240, 88, 370, 101 ], "spans": [ { "bbox": [ 240, 88, 370, 101 ], "score": 1.0, "content": "Table 9: Generator architecture.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 108, 105, 501, 394 ], "group_id": 0, "lines": [ { "bbox": [ 108, 105, 501, 394 ], "spans": [ { "bbox": [ 108, 105, 501, 394 ], "score": 0.979, "html": "
(a)Encoder
LayerResampleNormOutput Shape
Image x-128 × 128×3
Conv 1×1-128 ×128 ×128
ResBlkAvgPoolIN64× 64×256
ResBlkAvgPoolIN32 × 32 × 512
ResBlkAvgPoolIN16 × 16 × 512
ResBlkIN16 ×16× 512
ResBlkIN16 ×16 × 512
(b) Background Decoder(c) Foreground Decoder
LayerResampleNorm Output ShapeLayerResample NormOutput Shape
Input-16 × 16 × 448Input ResBlk16 × 16 × 64
ResBlkIN 16 ×16× 448 16 ×16× 512ResBlkAdaIN16 ×16× 64
ResBlkIN=AdaIN16 × 16 × 256
ResBlk=IN16 ×16 × 512ResBlkAdaIN16 ×16× 256
ResBlkUpsampleIN32 × 32×512ResBlk Upsample ResBlkAdaIN32 × 32 × 256
ResBlk ResBlkUpsample UpsampleIN IN64 × 64× 256 128 × 128× 448UpsampleAdaIN64×64×128
ResBlkUpsampleAdaIN128 ×128 × 64
(d) Fusion
LayerResampleNormOutput Shape
Input128 × 128 × (448 + 64)
Conv 1×1128 × 128× 3
", "type": "table", "image_path": "1c48ba4f6a9c5f47866cc9bb38143dabaa253479bdaae65b20f932c892e1f632.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 108, 105, 501, 201.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 108, 201.33333333333331, 501, 297.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 108, 297.66666666666663, 501, 393.99999999999994 ], "spans": [], "index": 3 } ] } ], "index": 1.0 }, { "type": "table", "bbox": [ 106, 439, 505, 637 ], "blocks": [ { "type": "table_caption", "bbox": [ 224, 421, 386, 432 ], "group_id": 1, "lines": [ { "bbox": [ 223, 420, 388, 433 ], "spans": [ { "bbox": [ 223, 420, 388, 433 ], "score": 1.0, "content": "Table 10: Mapping network architecture.", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "table_body", "bbox": [ 106, 439, 505, 637 ], "group_id": 1, "lines": [ { "bbox": [ 106, 439, 505, 637 ], "spans": [ { "bbox": [ 106, 439, 505, 637 ], "score": 0.966, "html": "
LayerActivationOutput Shape
Latent z=16
LinearReLU512
LinearReLU512
LinearReLU512
LinearReLU512
(b) Style Code(c) Content
LayerActivation Output ShapeLayerResample ActivationNoise Output Shape
Input=512Input=512
LinearReLU512Reshape--1×1×512
LinearReLU512ResBlkUpsampleINTrue2×2×512
LinearReLU512ResBlkUpsampleINTrue4×4×512
Linear164ResBlkUpsampleINTrue8×8×256
ResBlkUpsampleINTrue16 ×16×128
Conv 1×1=INTrue16 × 16× 64
", "type": "table", "image_path": "1098199c42bf1a6e4062025eaaa671f3e1525ccdf2528cdf76e0b9528b15881a.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 106, 439, 505, 505.0 ], "spans": [], "index": 5 }, { "bbox": [ 106, 505.0, 505, 571.0 ], "spans": [], "index": 6 }, { "bbox": [ 106, 571.0, 505, 637.0 ], "spans": [], "index": 7 } ] } ], "index": 5.0 }, { "type": "text", "bbox": [ 106, 665, 505, 732 ], "lines": [ { "bbox": [ 106, 666, 505, 678 ], "spans": [ { "bbox": [ 106, 666, 505, 678 ], "score": 1.0, "content": "Generator (Table 9). For the SDI dataset, the encoder part of the generator consists of three down-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "sampling blocks and two intermediate blocks (Table 9 (a)), all of them are pre-activation residual", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 687, 505, 700 ], "spans": [ { "bbox": [ 106, 687, 505, 700 ], "score": 1.0, "content": "units (He et al., 2016b). Then the encoded feature map is split channel-wise into background (Ta-", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 698, 506, 711 ], "spans": [ { "bbox": [ 106, 698, 506, 711 ], "score": 1.0, "content": "ble 9 (b)) and foreground (Table 9 (c)). Both of them are then carried through separate decoders. We", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 710, 505, 721 ], "spans": [ { "bbox": [ 106, 710, 505, 721 ], "score": 1.0, "content": "use the instance normalization (IN) and the adaptive instance normalization (AdaIN) as indicated.", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 721, 506, 733 ], "spans": [ { "bbox": [ 106, 721, 506, 733 ], "score": 1.0, "content": "The style code is injected into all AdaIN layers to modulate the affine transformations. Note that", "type": "text" } ], "index": 13 } ], "index": 10.5, "bbox_fs": [ 105, 666, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 105, 104, 511, 320 ], "blocks": [ { "type": "table_caption", "bbox": [ 176, 90, 434, 101 ], "group_id": 0, "lines": [ { "bbox": [ 177, 89, 435, 102 ], "spans": [ { "bbox": [ 177, 89, 435, 102 ], "score": 1.0, "content": "Table 11: Style-content encoder and discriminator architectures.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 105, 104, 511, 320 ], "group_id": 0, "lines": [ { "bbox": [ 105, 104, 511, 320 ], "spans": [ { "bbox": [ 105, 104, 511, 320 ], "score": 0.958, "html": "
(a)SharedLayers
LayerResampleNormOutput Shape
Input x128 × 128 ×3
Conv 1×11128 × 128 × 64
ResBlkAvgPool64 × 64× 256
ResBlkAvgPool32 × 32×512
ResBlkAvgPool16 ×16 × 512
(b) Style Code /Discriminator and BG Classifier(c) Content /FG Classifier
LayerResample NormOutput ShapeLayerResample NormOutput Shape
Input-16 ×16× 512Input16 ×16× 512
ResBlkAvgPool8×8×512LReLU16 ×16 × 512
ResBlkAvgPool4×4×512Conv 1×1*K16×16×64*K
LReLU4×4×512
Conv 4×41×1×512
LReLU1×1×512
Reshape512
Linear *KD*K
", "type": "table", "image_path": "0490c17ed37be089529218dd050adf14532c06146d2b840c7e6b071380cd65e8.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 105, 104, 511, 176.0 ], "spans": [], "index": 1 }, { "bbox": [ 105, 176.0, 511, 248.0 ], "spans": [], "index": 2 }, { "bbox": [ 105, 248.0, 511, 320.0 ], "spans": [], "index": 3 } ] } ], "index": 1.0 }, { "type": "text", "bbox": [ 107, 339, 504, 362 ], "lines": [ { "bbox": [ 105, 338, 505, 352 ], "spans": [ { "bbox": [ 105, 338, 505, 352 ], "score": 1.0, "content": "AdaIN is only used in the foreground decoder. The outputs of both decoders are only fused in the", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 349, 178, 362 ], "spans": [ { "bbox": [ 105, 349, 178, 362 ], "score": 1.0, "content": "end (Table 9 (d)).", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "text", "bbox": [ 106, 367, 505, 444 ], "lines": [ { "bbox": [ 106, 367, 505, 379 ], "spans": [ { "bbox": [ 106, 367, 505, 379 ], "score": 1.0, "content": "Mapping Network (Table 10). The mapping network consists of four shared linear layers (Table 10", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 378, 506, 391 ], "spans": [ { "bbox": [ 105, 378, 506, 391 ], "score": 1.0, "content": "(a)) and two separate branches: one for generating style codes (Table 10(b)) and one for contents", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 389, 505, 400 ], "spans": [ { "bbox": [ 105, 389, 307, 400 ], "score": 1.0, "content": "(Table 10(c)). Each of them is further divided into", "type": "text" }, { "bbox": [ 307, 389, 318, 399 ], "score": 0.81, "content": "K", "type": "inline_equation" }, { "bbox": [ 318, 389, 413, 400 ], "score": 1.0, "content": "output branches, where", "type": "text" }, { "bbox": [ 414, 389, 424, 399 ], "score": 0.78, "content": "K", "type": "inline_equation" }, { "bbox": [ 424, 389, 505, 400 ], "score": 1.0, "content": "denotes the number", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 401, 504, 412 ], "spans": [ { "bbox": [ 106, 401, 504, 412 ], "score": 1.0, "content": "of domains. The dimension of the input, the output style code and the output content is set to 16, 64,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 411, 505, 423 ], "spans": [ { "bbox": [ 105, 411, 124, 423 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 124, 411, 181, 422 ], "score": 0.92, "content": "1 6 \\times 1 6 \\times 6 4", "type": "inline_equation" }, { "bbox": [ 182, 411, 505, 423 ], "score": 1.0, "content": ", respectively. The latent code is sampled from the standard normal distribution.", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 421, 505, 434 ], "spans": [ { "bbox": [ 106, 421, 505, 434 ], "score": 1.0, "content": "Note that we apply per-pixel noise after each convolution in the content branch, which we have", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 432, 448, 446 ], "spans": [ { "bbox": [ 105, 432, 448, 446 ], "score": 1.0, "content": "observed to increase the diversity of generated defects significantly (cf. Figure 5 (g)).", "type": "text" } ], "index": 12 } ], "index": 9 }, { "type": "text", "bbox": [ 107, 450, 505, 505 ], "lines": [ { "bbox": [ 105, 448, 505, 462 ], "spans": [ { "bbox": [ 105, 448, 505, 462 ], "score": 1.0, "content": "Style-Content Encoder (Table 11). The style-content encoder consists of a CNN (Table 11 (a)) with", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 460, 505, 474 ], "spans": [ { "bbox": [ 105, 460, 432, 474 ], "score": 1.0, "content": "two branches (Table 11 (b) and (c)) as in the mapping network. Each branch has", "type": "text" }, { "bbox": [ 432, 461, 442, 471 ], "score": 0.82, "content": "K", "type": "inline_equation" }, { "bbox": [ 443, 460, 505, 474 ], "score": 1.0, "content": "outputs, where", "type": "text" } ], "index": 14 }, { "bbox": [ 107, 471, 506, 485 ], "spans": [ { "bbox": [ 107, 472, 117, 482 ], "score": 0.77, "content": "K", "type": "inline_equation" }, { "bbox": [ 118, 471, 506, 485 ], "score": 1.0, "content": "is the number of domains. Three pre-activation residual blocks are shared among two branches,", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 482, 506, 495 ], "spans": [ { "bbox": [ 105, 482, 396, 495 ], "score": 1.0, "content": "followed by a specific structure for each branch. The output dimension", "type": "text" }, { "bbox": [ 397, 483, 406, 492 ], "score": 0.78, "content": "D", "type": "inline_equation" }, { "bbox": [ 407, 482, 506, 495 ], "score": 1.0, "content": "in Table 11 is set to 64,", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 493, 294, 505 ], "spans": [ { "bbox": [ 106, 493, 294, 505 ], "score": 1.0, "content": "which denotes the dimension of the style code.", "type": "text" } ], "index": 17 } ], "index": 15 }, { "type": "text", "bbox": [ 106, 510, 505, 577 ], "lines": [ { "bbox": [ 106, 511, 504, 522 ], "spans": [ { "bbox": [ 106, 511, 504, 522 ], "score": 1.0, "content": "Discriminator (Table 11). The discriminator is a multi-task discriminator with two auxiliary clas-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 521, 505, 534 ], "spans": [ { "bbox": [ 106, 521, 505, 534 ], "score": 1.0, "content": "sifiers for the foreground content and the background. The structure is almost identical to the style-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 532, 505, 545 ], "spans": [ { "bbox": [ 105, 532, 204, 545 ], "score": 1.0, "content": "content encoder, except", "type": "text" }, { "bbox": [ 204, 533, 214, 542 ], "score": 0.77, "content": "D", "type": "inline_equation" }, { "bbox": [ 214, 532, 505, 545 ], "score": 1.0, "content": "is set to 1 for real/fake classification. The background classifier acts in", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 542, 505, 556 ], "spans": [ { "bbox": [ 104, 542, 505, 556 ], "score": 1.0, "content": "parallel to final linear layer in Table 11 (b) and provides the logits for background classification. The", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 555, 505, 566 ], "spans": [ { "bbox": [ 106, 555, 505, 566 ], "score": 1.0, "content": "foreground classifier instead acts on top of the output in Table 11 (c) and four more pre-activation", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 566, 493, 577 ], "spans": [ { "bbox": [ 106, 566, 493, 577 ], "score": 1.0, "content": "residual layers are applied to encode the content into logits for foreground content classification.", "type": "text" } ], "index": 23 } ], "index": 20.5 }, { "type": "title", "bbox": [ 107, 592, 244, 605 ], "lines": [ { "bbox": [ 105, 591, 245, 607 ], "spans": [ { "bbox": [ 105, 591, 245, 607 ], "score": 1.0, "content": "E ADDITIONAL RESULTS", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "title", "bbox": [ 108, 617, 323, 628 ], "lines": [ { "bbox": [ 106, 617, 325, 630 ], "spans": [ { "bbox": [ 106, 617, 325, 630 ], "score": 1.0, "content": "E.1 ADDITIONAL RESULTS ON THE SDI DATASET", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 637, 505, 704 ], "lines": [ { "bbox": [ 106, 637, 505, 650 ], "spans": [ { "bbox": [ 106, 637, 505, 650 ], "score": 1.0, "content": "We provide additional reference-guided image synthesis results on the SDI dataset in Figure 10.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 648, 505, 661 ], "spans": [ { "bbox": [ 105, 648, 505, 661 ], "score": 1.0, "content": "We demonstrate all the possible transfers among all foreground domains. Both style codes and", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "contents are extracted from the reference images. To be noted that DT-GAN can append and remove", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 671, 505, 684 ], "spans": [ { "bbox": [ 106, 671, 505, 684 ], "score": 1.0, "content": "foreground defects not only onto Normal samples but also to defective samples. For example,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 681, 505, 694 ], "spans": [ { "bbox": [ 105, 681, 505, 694 ], "score": 1.0, "content": "in the fifth column of Figure 10, the original scratch in the source image is removed and only the", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 691, 385, 706 ], "spans": [ { "bbox": [ 105, 691, 385, 706 ], "score": 1.0, "content": "defects from the reference images are presented in the output images.", "type": "text" } ], "index": 31 } ], "index": 28.5 }, { "type": "text", "bbox": [ 106, 709, 504, 732 ], "lines": [ { "bbox": [ 105, 708, 505, 723 ], "spans": [ { "bbox": [ 105, 708, 505, 723 ], "score": 1.0, "content": "Besides, we present additional evaluations showing the effectiveness of our synthetic data according", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 720, 505, 733 ], "spans": [ { "bbox": [ 105, 720, 505, 733 ], "score": 1.0, "content": "to Table 3. As seen in Table 12, the synthetic images from DT-GAN also boost the performance in", "type": "text" } ], "index": 33 } ], "index": 32.5 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 105, 104, 511, 320 ], "blocks": [ { "type": "table_caption", "bbox": [ 176, 90, 434, 101 ], "group_id": 0, "lines": [ { "bbox": [ 177, 89, 435, 102 ], "spans": [ { "bbox": [ 177, 89, 435, 102 ], "score": 1.0, "content": "Table 11: Style-content encoder and discriminator architectures.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 105, 104, 511, 320 ], "group_id": 0, "lines": [ { "bbox": [ 105, 104, 511, 320 ], "spans": [ { "bbox": [ 105, 104, 511, 320 ], "score": 0.958, "html": "
(a)SharedLayers
LayerResampleNormOutput Shape
Input x128 × 128 ×3
Conv 1×11128 × 128 × 64
ResBlkAvgPool64 × 64× 256
ResBlkAvgPool32 × 32×512
ResBlkAvgPool16 ×16 × 512
(b) Style Code /Discriminator and BG Classifier(c) Content /FG Classifier
LayerResample NormOutput ShapeLayerResample NormOutput Shape
Input-16 ×16× 512Input16 ×16× 512
ResBlkAvgPool8×8×512LReLU16 ×16 × 512
ResBlkAvgPool4×4×512Conv 1×1*K16×16×64*K
LReLU4×4×512
Conv 4×41×1×512
LReLU1×1×512
Reshape512
Linear *KD*K
", "type": "table", "image_path": "0490c17ed37be089529218dd050adf14532c06146d2b840c7e6b071380cd65e8.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 105, 104, 511, 176.0 ], "spans": [], "index": 1 }, { "bbox": [ 105, 176.0, 511, 248.0 ], "spans": [], "index": 2 }, { "bbox": [ 105, 248.0, 511, 320.0 ], "spans": [], "index": 3 } ] } ], "index": 1.0 }, { "type": "text", "bbox": [ 107, 339, 504, 362 ], "lines": [ { "bbox": [ 105, 338, 505, 352 ], "spans": [ { "bbox": [ 105, 338, 505, 352 ], "score": 1.0, "content": "AdaIN is only used in the foreground decoder. The outputs of both decoders are only fused in the", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 349, 178, 362 ], "spans": [ { "bbox": [ 105, 349, 178, 362 ], "score": 1.0, "content": "end (Table 9 (d)).", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 105, 338, 505, 362 ] }, { "type": "text", "bbox": [ 106, 367, 505, 444 ], "lines": [ { "bbox": [ 106, 367, 505, 379 ], "spans": [ { "bbox": [ 106, 367, 505, 379 ], "score": 1.0, "content": "Mapping Network (Table 10). The mapping network consists of four shared linear layers (Table 10", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 378, 506, 391 ], "spans": [ { "bbox": [ 105, 378, 506, 391 ], "score": 1.0, "content": "(a)) and two separate branches: one for generating style codes (Table 10(b)) and one for contents", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 389, 505, 400 ], "spans": [ { "bbox": [ 105, 389, 307, 400 ], "score": 1.0, "content": "(Table 10(c)). Each of them is further divided into", "type": "text" }, { "bbox": [ 307, 389, 318, 399 ], "score": 0.81, "content": "K", "type": "inline_equation" }, { "bbox": [ 318, 389, 413, 400 ], "score": 1.0, "content": "output branches, where", "type": "text" }, { "bbox": [ 414, 389, 424, 399 ], "score": 0.78, "content": "K", "type": "inline_equation" }, { "bbox": [ 424, 389, 505, 400 ], "score": 1.0, "content": "denotes the number", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 401, 504, 412 ], "spans": [ { "bbox": [ 106, 401, 504, 412 ], "score": 1.0, "content": "of domains. The dimension of the input, the output style code and the output content is set to 16, 64,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 411, 505, 423 ], "spans": [ { "bbox": [ 105, 411, 124, 423 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 124, 411, 181, 422 ], "score": 0.92, "content": "1 6 \\times 1 6 \\times 6 4", "type": "inline_equation" }, { "bbox": [ 182, 411, 505, 423 ], "score": 1.0, "content": ", respectively. The latent code is sampled from the standard normal distribution.", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 421, 505, 434 ], "spans": [ { "bbox": [ 106, 421, 505, 434 ], "score": 1.0, "content": "Note that we apply per-pixel noise after each convolution in the content branch, which we have", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 432, 448, 446 ], "spans": [ { "bbox": [ 105, 432, 448, 446 ], "score": 1.0, "content": "observed to increase the diversity of generated defects significantly (cf. Figure 5 (g)).", "type": "text" } ], "index": 12 } ], "index": 9, "bbox_fs": [ 105, 367, 506, 446 ] }, { "type": "text", "bbox": [ 107, 450, 505, 505 ], "lines": [ { "bbox": [ 105, 448, 505, 462 ], "spans": [ { "bbox": [ 105, 448, 505, 462 ], "score": 1.0, "content": "Style-Content Encoder (Table 11). The style-content encoder consists of a CNN (Table 11 (a)) with", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 460, 505, 474 ], "spans": [ { "bbox": [ 105, 460, 432, 474 ], "score": 1.0, "content": "two branches (Table 11 (b) and (c)) as in the mapping network. Each branch has", "type": "text" }, { "bbox": [ 432, 461, 442, 471 ], "score": 0.82, "content": "K", "type": "inline_equation" }, { "bbox": [ 443, 460, 505, 474 ], "score": 1.0, "content": "outputs, where", "type": "text" } ], "index": 14 }, { "bbox": [ 107, 471, 506, 485 ], "spans": [ { "bbox": [ 107, 472, 117, 482 ], "score": 0.77, "content": "K", "type": "inline_equation" }, { "bbox": [ 118, 471, 506, 485 ], "score": 1.0, "content": "is the number of domains. Three pre-activation residual blocks are shared among two branches,", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 482, 506, 495 ], "spans": [ { "bbox": [ 105, 482, 396, 495 ], "score": 1.0, "content": "followed by a specific structure for each branch. The output dimension", "type": "text" }, { "bbox": [ 397, 483, 406, 492 ], "score": 0.78, "content": "D", "type": "inline_equation" }, { "bbox": [ 407, 482, 506, 495 ], "score": 1.0, "content": "in Table 11 is set to 64,", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 493, 294, 505 ], "spans": [ { "bbox": [ 106, 493, 294, 505 ], "score": 1.0, "content": "which denotes the dimension of the style code.", "type": "text" } ], "index": 17 } ], "index": 15, "bbox_fs": [ 105, 448, 506, 505 ] }, { "type": "text", "bbox": [ 106, 510, 505, 577 ], "lines": [ { "bbox": [ 106, 511, 504, 522 ], "spans": [ { "bbox": [ 106, 511, 504, 522 ], "score": 1.0, "content": "Discriminator (Table 11). The discriminator is a multi-task discriminator with two auxiliary clas-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 521, 505, 534 ], "spans": [ { "bbox": [ 106, 521, 505, 534 ], "score": 1.0, "content": "sifiers for the foreground content and the background. The structure is almost identical to the style-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 532, 505, 545 ], "spans": [ { "bbox": [ 105, 532, 204, 545 ], "score": 1.0, "content": "content encoder, except", "type": "text" }, { "bbox": [ 204, 533, 214, 542 ], "score": 0.77, "content": "D", "type": "inline_equation" }, { "bbox": [ 214, 532, 505, 545 ], "score": 1.0, "content": "is set to 1 for real/fake classification. The background classifier acts in", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 542, 505, 556 ], "spans": [ { "bbox": [ 104, 542, 505, 556 ], "score": 1.0, "content": "parallel to final linear layer in Table 11 (b) and provides the logits for background classification. The", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 555, 505, 566 ], "spans": [ { "bbox": [ 106, 555, 505, 566 ], "score": 1.0, "content": "foreground classifier instead acts on top of the output in Table 11 (c) and four more pre-activation", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 566, 493, 577 ], "spans": [ { "bbox": [ 106, 566, 493, 577 ], "score": 1.0, "content": "residual layers are applied to encode the content into logits for foreground content classification.", "type": "text" } ], "index": 23 } ], "index": 20.5, "bbox_fs": [ 104, 511, 505, 577 ] }, { "type": "title", "bbox": [ 107, 592, 244, 605 ], "lines": [ { "bbox": [ 105, 591, 245, 607 ], "spans": [ { "bbox": [ 105, 591, 245, 607 ], "score": 1.0, "content": "E ADDITIONAL RESULTS", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "title", "bbox": [ 108, 617, 323, 628 ], "lines": [ { "bbox": [ 106, 617, 325, 630 ], "spans": [ { "bbox": [ 106, 617, 325, 630 ], "score": 1.0, "content": "E.1 ADDITIONAL RESULTS ON THE SDI DATASET", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 637, 505, 704 ], "lines": [ { "bbox": [ 106, 637, 505, 650 ], "spans": [ { "bbox": [ 106, 637, 505, 650 ], "score": 1.0, "content": "We provide additional reference-guided image synthesis results on the SDI dataset in Figure 10.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 648, 505, 661 ], "spans": [ { "bbox": [ 105, 648, 505, 661 ], "score": 1.0, "content": "We demonstrate all the possible transfers among all foreground domains. Both style codes and", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 659, 506, 673 ], "spans": [ { "bbox": [ 105, 659, 506, 673 ], "score": 1.0, "content": "contents are extracted from the reference images. To be noted that DT-GAN can append and remove", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 671, 505, 684 ], "spans": [ { "bbox": [ 106, 671, 505, 684 ], "score": 1.0, "content": "foreground defects not only onto Normal samples but also to defective samples. For example,", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 681, 505, 694 ], "spans": [ { "bbox": [ 105, 681, 505, 694 ], "score": 1.0, "content": "in the fifth column of Figure 10, the original scratch in the source image is removed and only the", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 691, 385, 706 ], "spans": [ { "bbox": [ 105, 691, 385, 706 ], "score": 1.0, "content": "defects from the reference images are presented in the output images.", "type": "text" } ], "index": 31 } ], "index": 28.5, "bbox_fs": [ 105, 637, 506, 706 ] }, { "type": "text", "bbox": [ 106, 709, 504, 732 ], "lines": [ { "bbox": [ 105, 708, 505, 723 ], "spans": [ { "bbox": [ 105, 708, 505, 723 ], "score": 1.0, "content": "Besides, we present additional evaluations showing the effectiveness of our synthetic data according", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 720, 505, 733 ], "spans": [ { "bbox": [ 105, 720, 505, 733 ], "score": 1.0, "content": "to Table 3. As seen in Table 12, the synthetic images from DT-GAN also boost the performance in", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 582, 505, 596 ], "spans": [ { "bbox": [ 105, 582, 505, 596 ], "score": 1.0, "content": "single product classifiers, where the classifiers were trained on the subset of images for one product", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 594, 281, 606 ], "spans": [ { "bbox": [ 105, 594, 281, 606 ], "score": 1.0, "content": "(A, B, C) instead of the full dataset (ABC).", "type": "text", "cross_page": true } ], "index": 9 } ], "index": 32.5, "bbox_fs": [ 105, 708, 505, 733 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 109, 78, 504, 483 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 78, 504, 483 ], "group_id": 0, "lines": [ { "bbox": [ 109, 78, 504, 483 ], "spans": [ { "bbox": [ 109, 78, 504, 483 ], "score": 0.959, "type": "image", "image_path": "ba0e2234e87b25b63cc95ac4e5644f8da2a00a1e319667a7cf9fa0a32eed56c7.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 78, 504, 213.0 ], "spans": [], "index": 0 }, { "bbox": [ 109, 213.0, 504, 348.0 ], "spans": [], "index": 1 }, { "bbox": [ 109, 348.0, 504, 483.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 493, 505, 549 ], "group_id": 0, "lines": [ { "bbox": [ 106, 494, 505, 506 ], "spans": [ { "bbox": [ 106, 494, 505, 506 ], "score": 1.0, "content": "Figure 10: Reference-guided image synthesis results on the SDI dataset. The first row and the first", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 505, 505, 517 ], "spans": [ { "bbox": [ 106, 505, 505, 517 ], "score": 1.0, "content": "column are the real images sampled from the dataset, while the rest are synthetic images generated", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 516, 505, 528 ], "spans": [ { "bbox": [ 106, 516, 505, 528 ], "score": 1.0, "content": "by the proposed DT-GAN. Our model provides translations between different foreground domains", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 525, 506, 541 ], "spans": [ { "bbox": [ 105, 525, 506, 541 ], "score": 1.0, "content": "(Normal, Scratches and Spots) with styles and contents extracted from reference images", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 537, 358, 551 ], "spans": [ { "bbox": [ 105, 537, 358, 551 ], "score": 1.0, "content": "while the backgrounds from source images are well preserved.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 582, 504, 605 ], "lines": [ { "bbox": [ 105, 582, 505, 596 ], "spans": [ { "bbox": [ 105, 582, 505, 596 ], "score": 1.0, "content": "single product classifiers, where the classifiers were trained on the subset of images for one product", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 594, 281, 606 ], "spans": [ { "bbox": [ 105, 594, 281, 606 ], "score": 1.0, "content": "(A, B, C) instead of the full dataset (ABC).", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "text", "bbox": [ 106, 611, 505, 732 ], "lines": [ { "bbox": [ 105, 610, 505, 623 ], "spans": [ { "bbox": [ 105, 610, 505, 623 ], "score": 1.0, "content": "As discussed in Section 4.2, we assumed that the data-insufficiency problem can be mitigated by", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 622, 505, 635 ], "spans": [ { "bbox": [ 105, 622, 505, 635 ], "score": 1.0, "content": "transferring defects across multiple background products. To examine if this assumption holds, we", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 632, 505, 646 ], "spans": [ { "bbox": [ 105, 632, 505, 646 ], "score": 1.0, "content": "compared the performance of classifiers trained on synthetic images with defects from a specific", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 644, 505, 656 ], "spans": [ { "bbox": [ 105, 644, 505, 656 ], "score": 1.0, "content": "source (vA, vB, vC) to classifiers trained on images with defects from all products (vABC). The", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 655, 505, 667 ], "spans": [ { "bbox": [ 105, 655, 505, 667 ], "score": 1.0, "content": "results on the cross-domain effect with regard to different sizes of the training set are shown in", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 664, 506, 679 ], "spans": [ { "bbox": [ 105, 664, 506, 679 ], "score": 1.0, "content": "Table 13. We again notice that using our synthetic data is beneficial. Moreover, in most cases the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "performance is further improved by exploiting cross-domain information (i.e. by transferring defects", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 688, 506, 700 ], "spans": [ { "bbox": [ 106, 688, 506, 700 ], "score": 1.0, "content": "from other products). We interpret this as support for our assumption and the practical usefulness of", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "our method in the real-world scenario. The case of cross-domain image synthesis when the desired", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "combination is not presented in the training set is covered in the study on the MVTec Anomaly", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 721, 357, 733 ], "spans": [ { "bbox": [ 106, 721, 357, 733 ], "score": 1.0, "content": "Detection dataset (Bergmann et al., 2019) (see Appendix E.4).", "type": "text" } ], "index": 20 } ], "index": 15 } ], "page_idx": 17, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 763 ], "spans": [ { "bbox": [ 299, 750, 312, 763 ], "score": 1.0, "content": "18", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 309, 39 ], "spans": [ { "bbox": [ 106, 25, 309, 39 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 109, 78, 504, 483 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 78, 504, 483 ], "group_id": 0, "lines": [ { "bbox": [ 109, 78, 504, 483 ], "spans": [ { "bbox": [ 109, 78, 504, 483 ], "score": 0.959, "type": "image", "image_path": "ba0e2234e87b25b63cc95ac4e5644f8da2a00a1e319667a7cf9fa0a32eed56c7.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 78, 504, 213.0 ], "spans": [], "index": 0 }, { "bbox": [ 109, 213.0, 504, 348.0 ], "spans": [], "index": 1 }, { "bbox": [ 109, 348.0, 504, 483.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 493, 505, 549 ], "group_id": 0, "lines": [ { "bbox": [ 106, 494, 505, 506 ], "spans": [ { "bbox": [ 106, 494, 505, 506 ], "score": 1.0, "content": "Figure 10: Reference-guided image synthesis results on the SDI dataset. The first row and the first", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 505, 505, 517 ], "spans": [ { "bbox": [ 106, 505, 505, 517 ], "score": 1.0, "content": "column are the real images sampled from the dataset, while the rest are synthetic images generated", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 516, 505, 528 ], "spans": [ { "bbox": [ 106, 516, 505, 528 ], "score": 1.0, "content": "by the proposed DT-GAN. Our model provides translations between different foreground domains", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 525, 506, 541 ], "spans": [ { "bbox": [ 105, 525, 506, 541 ], "score": 1.0, "content": "(Normal, Scratches and Spots) with styles and contents extracted from reference images", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 537, 358, 551 ], "spans": [ { "bbox": [ 105, 537, 358, 551 ], "score": 1.0, "content": "while the backgrounds from source images are well preserved.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 }, { "type": "text", "bbox": [ 107, 582, 504, 605 ], "lines": [], "index": 8.5, "bbox_fs": [ 105, 582, 505, 606 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 611, 505, 732 ], "lines": [ { "bbox": [ 105, 610, 505, 623 ], "spans": [ { "bbox": [ 105, 610, 505, 623 ], "score": 1.0, "content": "As discussed in Section 4.2, we assumed that the data-insufficiency problem can be mitigated by", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 622, 505, 635 ], "spans": [ { "bbox": [ 105, 622, 505, 635 ], "score": 1.0, "content": "transferring defects across multiple background products. To examine if this assumption holds, we", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 632, 505, 646 ], "spans": [ { "bbox": [ 105, 632, 505, 646 ], "score": 1.0, "content": "compared the performance of classifiers trained on synthetic images with defects from a specific", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 644, 505, 656 ], "spans": [ { "bbox": [ 105, 644, 505, 656 ], "score": 1.0, "content": "source (vA, vB, vC) to classifiers trained on images with defects from all products (vABC). The", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 655, 505, 667 ], "spans": [ { "bbox": [ 105, 655, 505, 667 ], "score": 1.0, "content": "results on the cross-domain effect with regard to different sizes of the training set are shown in", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 664, 506, 679 ], "spans": [ { "bbox": [ 105, 664, 506, 679 ], "score": 1.0, "content": "Table 13. We again notice that using our synthetic data is beneficial. Moreover, in most cases the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "performance is further improved by exploiting cross-domain information (i.e. by transferring defects", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 688, 506, 700 ], "spans": [ { "bbox": [ 106, 688, 506, 700 ], "score": 1.0, "content": "from other products). We interpret this as support for our assumption and the practical usefulness of", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "our method in the real-world scenario. The case of cross-domain image synthesis when the desired", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "combination is not presented in the training set is covered in the study on the MVTec Anomaly", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 721, 357, 733 ], "spans": [ { "bbox": [ 106, 721, 357, 733 ], "score": 1.0, "content": "Detection dataset (Bergmann et al., 2019) (see Appendix E.4).", "type": "text" } ], "index": 20 } ], "index": 15, "bbox_fs": [ 105, 610, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 109, 134, 500, 323 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 89, 505, 133 ], "group_id": 0, "lines": [ { "bbox": [ 106, 89, 505, 101 ], "spans": [ { "bbox": [ 106, 89, 505, 101 ], "score": 1.0, "content": "Table 12: Quantitative results for DT-GAN as a data augmentation method to train general and single", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 100, 505, 112 ], "spans": [ { "bbox": [ 105, 100, 505, 112 ], "score": 1.0, "content": "product classifiers. The left-most column indicates the number of samples per class, including all", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 111, 505, 123 ], "spans": [ { "bbox": [ 106, 111, 505, 123 ], "score": 1.0, "content": "images from the training set plus increasing amounts of synthetic images. In the first row, 20A refers", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 122, 479, 134 ], "spans": [ { "bbox": [ 105, 122, 479, 134 ], "score": 1.0, "content": "to the case of 20 real defective samples for product A, while All refers to the full training set.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 109, 134, 500, 323 ], "group_id": 0, "lines": [ { "bbox": [ 109, 134, 500, 323 ], "spans": [ { "bbox": [ 109, 134, 500, 323 ], "score": 0.979, "html": "
Dataset Size20A
ABCABC
Trad-AugOursTrad-AugOursTrad-AugOursTrad-AugOurs
450035.09±2.62 27.64±3.127.8±1.485.6±1.6715.24±1.90 13.14±1.7015.55±0.63 14.28±1.25
660039.64±2.28 27.64±1.658.8±1.646.2±1.6415.81±1.73112.38±1.6516.69±0.76 14.41±3.12
1200034.18±4.39 28.55±7.325.8±0.455.6±1.1416.19±1.17 10.86±1.28 16.95±1.02 14.22±1.53
1860039.45±7.06 32.55±5.047.2±0.845.2±1.1014.86±0.85 13.14±2.06 16.12±2.19 15.36±0.86
Dataset SizeAll
ABCABC
Trad-AugOursTrad-AugOursTrad-AugOursTrad-AugOurs
450016.00±1.04110.18±1.75 8.79±0.455.60±1.5117.13±6.62 14.09±2.2712.75±0.61 11.04±0.76
660014.90±1.3810.54±1.22 7.60±1.516.80±3.1115.23±2.33 11.42±013.07±1.57 10.60±0.48
1200013.81±2.366.72±1.65 6.80±1.644.60±016.57±3.20 13.90±2.5712.05±0.819.90±0.69
1860013.63±2.22 10.54±2.45 6.80±1.794.99±1.87 15.62±0.85 11.61±1.24 12.37±0.32 10.21±0.96
", "type": "table", "image_path": "20974be996d9fce37f644ecdf9151a7e5df8554b4465f8cbe6a1c0ef5f25b6a6.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 109, 134, 500, 197.0 ], "spans": [], "index": 4 }, { "bbox": [ 109, 197.0, 500, 260.0 ], "spans": [], "index": 5 }, { "bbox": [ 109, 260.0, 500, 323.0 ], "spans": [], "index": 6 } ] } ], "index": 3.25 }, { "type": "table", "bbox": [ 147, 379, 461, 619 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 345, 504, 379 ], "group_id": 1, "lines": [ { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "Table 13: Cross-domain effect on single product classifiers trained with reference-guided synthetic", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 357, 506, 369 ], "spans": [ { "bbox": [ 106, 357, 506, 369 ], "score": 1.0, "content": "images at all scales. Note that here A, B and C stand for 3 products in the SDI dataset while vA, vB,", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 367, 394, 380 ], "spans": [ { "bbox": [ 105, 367, 394, 380 ], "score": 1.0, "content": "vC and vABC indicate the defects are copied from which reference set.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "table_body", "bbox": [ 147, 379, 461, 619 ], "group_id": 1, "lines": [ { "bbox": [ 147, 379, 461, 619 ], "spans": [ { "bbox": [ 147, 379, 461, 619 ], "score": 0.981, "html": "
Dataset SizeA
Trad-AugvAvBvCvABC
450016.00±1.0412.90±2.6113.08±1.6514.90±2.4615.27±3.49
660014.90±1.3813.99±1.8911.26±1.0414.36±4.0416.00±2.85
1200013.81±2.3611.81±2.6512.72±2.8711.99±1.6311.09±3.49
1860013.63±2.2212.72±5.2214.36±3.8314.18±5.0513.81±8.56
DatasetB
SizeTrad-AugvAvBvCVABC
45008.79±0.457.80±2.155.60±1.1410.19±0.846.79±1.30
66007.60±1.516.80±1.657.80±1.108.00±2.346.00±1.41
120006.80±1.646.40±1.346.60±1.526.59±1.345.60±1.34
186006.80±1.796.19±1.784.40±1.146.60±1.955.99±1.58
DatasetC
SizeTrad-AugvAVBvCvABC
450017.14±4.6214.85±0.5216.76±2.5813.90±1.9812.00±1.59
660015.23±2.3313.14±1.2413.90±2.2914.28±1.3412.57±1.57
1200016.57±3.2013.14±2.8111.23±0.8014.85±1.7311.42±0.96
1860015.62±0.8513.71±1.7315.99±6.7512.57±3.2612.95±2.98
", "type": "table", "image_path": "f17f93e8ebcc379ed4eada67e2de92805226db4f931acda640080a5aad9fc349.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 147, 379, 461, 459.0 ], "spans": [], "index": 10 }, { "bbox": [ 147, 459.0, 461, 539.0 ], "spans": [], "index": 11 }, { "bbox": [ 147, 539.0, 461, 619.0 ], "spans": [], "index": 12 } ] } ], "index": 9.5 }, { "type": "title", "bbox": [ 107, 642, 386, 655 ], "lines": [ { "bbox": [ 105, 642, 388, 656 ], "spans": [ { "bbox": [ 105, 642, 388, 656 ], "score": 1.0, "content": "E.2 ADDITIONAL FID AND KID RESULTS ON THE SDI DATASET", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 665, 505, 732 ], "lines": [ { "bbox": [ 106, 665, 505, 678 ], "spans": [ { "bbox": [ 106, 665, 505, 678 ], "score": 1.0, "content": "We provide additional results in the case of training GANs with augmentation methods in Table 14.", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 676, 506, 689 ], "spans": [ { "bbox": [ 105, 676, 506, 689 ], "score": 1.0, "content": "Augmentation methods like ADA (Karras et al., 2020a) or DiffAug (Zhao et al., 2020) are proposed", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 687, 506, 701 ], "spans": [ { "bbox": [ 105, 687, 506, 701 ], "score": 1.0, "content": "to adapt GAN training to limited data. We applied these augmentation methods to StyleGAN v2 and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "BigGAN, because these state-of-art image synthesis methods are not optimized for small dataset.", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "However, incorporating the augmentation methods in training GANs on the SDI dataset is not always", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 720, 505, 733 ], "spans": [ { "bbox": [ 105, 720, 505, 733 ], "score": 1.0, "content": "beneficial. The performance of StyleGAN v2 is largely degraded when using ADA, potentially due", "type": "text" } ], "index": 19 } ], "index": 16.5 } ], "page_idx": 18, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "19", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 109, 134, 500, 323 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 89, 505, 133 ], "group_id": 0, "lines": [ { "bbox": [ 106, 89, 505, 101 ], "spans": [ { "bbox": [ 106, 89, 505, 101 ], "score": 1.0, "content": "Table 12: Quantitative results for DT-GAN as a data augmentation method to train general and single", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 100, 505, 112 ], "spans": [ { "bbox": [ 105, 100, 505, 112 ], "score": 1.0, "content": "product classifiers. The left-most column indicates the number of samples per class, including all", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 111, 505, 123 ], "spans": [ { "bbox": [ 106, 111, 505, 123 ], "score": 1.0, "content": "images from the training set plus increasing amounts of synthetic images. In the first row, 20A refers", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 122, 479, 134 ], "spans": [ { "bbox": [ 105, 122, 479, 134 ], "score": 1.0, "content": "to the case of 20 real defective samples for product A, while All refers to the full training set.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 109, 134, 500, 323 ], "group_id": 0, "lines": [ { "bbox": [ 109, 134, 500, 323 ], "spans": [ { "bbox": [ 109, 134, 500, 323 ], "score": 0.979, "html": "
Dataset Size20A
ABCABC
Trad-AugOursTrad-AugOursTrad-AugOursTrad-AugOurs
450035.09±2.62 27.64±3.127.8±1.485.6±1.6715.24±1.90 13.14±1.7015.55±0.63 14.28±1.25
660039.64±2.28 27.64±1.658.8±1.646.2±1.6415.81±1.73112.38±1.6516.69±0.76 14.41±3.12
1200034.18±4.39 28.55±7.325.8±0.455.6±1.1416.19±1.17 10.86±1.28 16.95±1.02 14.22±1.53
1860039.45±7.06 32.55±5.047.2±0.845.2±1.1014.86±0.85 13.14±2.06 16.12±2.19 15.36±0.86
Dataset SizeAll
ABCABC
Trad-AugOursTrad-AugOursTrad-AugOursTrad-AugOurs
450016.00±1.04110.18±1.75 8.79±0.455.60±1.5117.13±6.62 14.09±2.2712.75±0.61 11.04±0.76
660014.90±1.3810.54±1.22 7.60±1.516.80±3.1115.23±2.33 11.42±013.07±1.57 10.60±0.48
1200013.81±2.366.72±1.65 6.80±1.644.60±016.57±3.20 13.90±2.5712.05±0.819.90±0.69
1860013.63±2.22 10.54±2.45 6.80±1.794.99±1.87 15.62±0.85 11.61±1.24 12.37±0.32 10.21±0.96
", "type": "table", "image_path": "20974be996d9fce37f644ecdf9151a7e5df8554b4465f8cbe6a1c0ef5f25b6a6.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 109, 134, 500, 197.0 ], "spans": [], "index": 4 }, { "bbox": [ 109, 197.0, 500, 260.0 ], "spans": [], "index": 5 }, { "bbox": [ 109, 260.0, 500, 323.0 ], "spans": [], "index": 6 } ] } ], "index": 3.25 }, { "type": "table", "bbox": [ 147, 379, 461, 619 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 345, 504, 379 ], "group_id": 1, "lines": [ { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "Table 13: Cross-domain effect on single product classifiers trained with reference-guided synthetic", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 357, 506, 369 ], "spans": [ { "bbox": [ 106, 357, 506, 369 ], "score": 1.0, "content": "images at all scales. Note that here A, B and C stand for 3 products in the SDI dataset while vA, vB,", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 367, 394, 380 ], "spans": [ { "bbox": [ 105, 367, 394, 380 ], "score": 1.0, "content": "vC and vABC indicate the defects are copied from which reference set.", "type": "text" } ], "index": 9 } ], "index": 8 }, { "type": "table_body", "bbox": [ 147, 379, 461, 619 ], "group_id": 1, "lines": [ { "bbox": [ 147, 379, 461, 619 ], "spans": [ { "bbox": [ 147, 379, 461, 619 ], "score": 0.981, "html": "
Dataset SizeA
Trad-AugvAvBvCvABC
450016.00±1.0412.90±2.6113.08±1.6514.90±2.4615.27±3.49
660014.90±1.3813.99±1.8911.26±1.0414.36±4.0416.00±2.85
1200013.81±2.3611.81±2.6512.72±2.8711.99±1.6311.09±3.49
1860013.63±2.2212.72±5.2214.36±3.8314.18±5.0513.81±8.56
DatasetB
SizeTrad-AugvAvBvCVABC
45008.79±0.457.80±2.155.60±1.1410.19±0.846.79±1.30
66007.60±1.516.80±1.657.80±1.108.00±2.346.00±1.41
120006.80±1.646.40±1.346.60±1.526.59±1.345.60±1.34
186006.80±1.796.19±1.784.40±1.146.60±1.955.99±1.58
DatasetC
SizeTrad-AugvAVBvCvABC
450017.14±4.6214.85±0.5216.76±2.5813.90±1.9812.00±1.59
660015.23±2.3313.14±1.2413.90±2.2914.28±1.3412.57±1.57
1200016.57±3.2013.14±2.8111.23±0.8014.85±1.7311.42±0.96
1860015.62±0.8513.71±1.7315.99±6.7512.57±3.2612.95±2.98
", "type": "table", "image_path": "f17f93e8ebcc379ed4eada67e2de92805226db4f931acda640080a5aad9fc349.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 147, 379, 461, 459.0 ], "spans": [], "index": 10 }, { "bbox": [ 147, 459.0, 461, 539.0 ], "spans": [], "index": 11 }, { "bbox": [ 147, 539.0, 461, 619.0 ], "spans": [], "index": 12 } ] } ], "index": 9.5 }, { "type": "title", "bbox": [ 107, 642, 386, 655 ], "lines": [ { "bbox": [ 105, 642, 388, 656 ], "spans": [ { "bbox": [ 105, 642, 388, 656 ], "score": 1.0, "content": "E.2 ADDITIONAL FID AND KID RESULTS ON THE SDI DATASET", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 665, 505, 732 ], "lines": [ { "bbox": [ 106, 665, 505, 678 ], "spans": [ { "bbox": [ 106, 665, 505, 678 ], "score": 1.0, "content": "We provide additional results in the case of training GANs with augmentation methods in Table 14.", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 676, 506, 689 ], "spans": [ { "bbox": [ 105, 676, 506, 689 ], "score": 1.0, "content": "Augmentation methods like ADA (Karras et al., 2020a) or DiffAug (Zhao et al., 2020) are proposed", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 687, 506, 701 ], "spans": [ { "bbox": [ 105, 687, 506, 701 ], "score": 1.0, "content": "to adapt GAN training to limited data. We applied these augmentation methods to StyleGAN v2 and", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "BigGAN, because these state-of-art image synthesis methods are not optimized for small dataset.", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "However, incorporating the augmentation methods in training GANs on the SDI dataset is not always", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 720, 505, 733 ], "spans": [ { "bbox": [ 105, 720, 505, 733 ], "score": 1.0, "content": "beneficial. The performance of StyleGAN v2 is largely degraded when using ADA, potentially due", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 83, 505, 95 ], "spans": [ { "bbox": [ 106, 83, 505, 95 ], "score": 1.0, "content": "to the conflict between augmentation methods and the decentralized location of defects—in the SDI", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 93, 505, 106 ], "spans": [ { "bbox": [ 106, 93, 505, 106 ], "score": 1.0, "content": "dataset, defects can occur anywhere on the surface. This is in contrast to datasets that were used", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 106, 105, 505, 117 ], "spans": [ { "bbox": [ 106, 105, 505, 117 ], "score": 1.0, "content": "to evaluate the aforementioned augmentation methods in GANs, where the objects are centralized", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 104, 113, 507, 131 ], "spans": [ { "bbox": [ 104, 113, 507, 131 ], "score": 1.0, "content": "(e.g., ImageNet (Deng et al., 2009), Cifar (Krizhevsky & Hinton, 2009)) and their attributes (e.g.,", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 106, 126, 439, 140 ], "spans": [ { "bbox": [ 106, 126, 439, 140 ], "score": 1.0, "content": "beard, eye glasses in CelebA (Liu et al., 2015)) only occur in specific images parts.", "type": "text", "cross_page": true } ], "index": 4 } ], "index": 16.5, "bbox_fs": [ 105, 665, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 138 ], "lines": [ { "bbox": [ 106, 83, 505, 95 ], "spans": [ { "bbox": [ 106, 83, 505, 95 ], "score": 1.0, "content": "to the conflict between augmentation methods and the decentralized location of defects—in the SDI", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 93, 505, 106 ], "spans": [ { "bbox": [ 106, 93, 505, 106 ], "score": 1.0, "content": "dataset, defects can occur anywhere on the surface. This is in contrast to datasets that were used", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 105, 505, 117 ], "spans": [ { "bbox": [ 106, 105, 505, 117 ], "score": 1.0, "content": "to evaluate the aforementioned augmentation methods in GANs, where the objects are centralized", "type": "text" } ], "index": 2 }, { "bbox": [ 104, 113, 507, 131 ], "spans": [ { "bbox": [ 104, 113, 507, 131 ], "score": 1.0, "content": "(e.g., ImageNet (Deng et al., 2009), Cifar (Krizhevsky & Hinton, 2009)) and their attributes (e.g.,", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 126, 439, 140 ], "spans": [ { "bbox": [ 106, 126, 439, 140 ], "score": 1.0, "content": "beard, eye glasses in CelebA (Liu et al., 2015)) only occur in specific images parts.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "table", "bbox": [ 106, 200, 503, 309 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 155, 505, 200 ], "group_id": 0, "lines": [ { "bbox": [ 105, 155, 506, 169 ], "spans": [ { "bbox": [ 105, 155, 506, 169 ], "score": 1.0, "content": "Table 14: Quantitative comparison of DT-GAN with baseline image synthesis methods using FID", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 167, 506, 180 ], "spans": [ { "bbox": [ 105, 167, 506, 180 ], "score": 1.0, "content": "and KID. Note that the reported values are not comparable between columns, because they are", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 177, 506, 190 ], "spans": [ { "bbox": [ 105, 177, 506, 190 ], "score": 1.0, "content": "calculated on different training sets. The scores of StarGAN v2 on single products are omitted", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 188, 488, 201 ], "spans": [ { "bbox": [ 105, 188, 488, 201 ], "score": 1.0, "content": "because generating images with specified background is not possible due to its network design.", "type": "text" } ], "index": 8 } ], "index": 6.5 }, { "type": "table_body", "bbox": [ 106, 200, 503, 309 ], "group_id": 0, "lines": [ { "bbox": [ 106, 200, 503, 309 ], "spans": [ { "bbox": [ 106, 200, 503, 309 ], "score": 0.982, "html": "
MethodFID↓KID↓
ABCAllABCAll
Mokady et al. (2020)68.6966.9036.2158.630.0500.0360.0300.036
StarGAN v2--137.701-10.013
StyleGAN v290.1052.95138.0935.340.0720.0270.1860.013
StyleGAN v2 + ADA149.6642.75135.6976.160.1380.0190.1910.055
BigGAN235.66192.89193.61151.430.2480.1990.2760.115
BigGAN + DiffAug218.74134.41270.89155.880.2200.1210.3780.099
Ours58.4336.4422.6829.730.0250.0130.0120.009
", "type": "table", "image_path": "45a92e9be03f105fca9e42ed49982648d41734c531d7e3f15ce1a34f30d32186.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 106, 200, 503, 236.33333333333334 ], "spans": [], "index": 9 }, { "bbox": [ 106, 236.33333333333334, 503, 272.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 106, 272.6666666666667, 503, 309.0 ], "spans": [], "index": 11 } ] } ], "index": 8.25 }, { "type": "title", "bbox": [ 107, 331, 385, 343 ], "lines": [ { "bbox": [ 105, 330, 386, 345 ], "spans": [ { "bbox": [ 105, 330, 386, 345 ], "score": 1.0, "content": "E.3 ABLATION STUDY WITH REGARD TO FID AND KID SCORES", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 351, 505, 451 ], "lines": [ { "bbox": [ 106, 352, 505, 365 ], "spans": [ { "bbox": [ 106, 352, 505, 365 ], "score": 1.0, "content": "We report the FID and KID scores of the ablation study in Table 15. We notice that both subnetworks", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 361, 506, 378 ], "spans": [ { "bbox": [ 104, 361, 506, 378 ], "score": 1.0, "content": "show positive correlation to each modification except for structural change as in (a) and (e) . Among", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 375, 505, 387 ], "spans": [ { "bbox": [ 106, 375, 505, 387 ], "score": 1.0, "content": "the two subnetworks, the reference-guided subnetwork outperforms the latent-guided one in the", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 385, 505, 397 ], "spans": [ { "bbox": [ 106, 385, 505, 397 ], "score": 1.0, "content": "beginning, which is due to the fact that transferring existing contents is easier than generating them", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 396, 505, 408 ], "spans": [ { "bbox": [ 106, 396, 505, 408 ], "score": 1.0, "content": "from random noise. This effect is also observed in Figure 5. However, the performance of the latent-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 407, 505, 420 ], "spans": [ { "bbox": [ 106, 407, 505, 420 ], "score": 1.0, "content": "guided subnetwork improves significantly after applying per-pixel noise injection. The subnetwork", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 419, 505, 430 ], "spans": [ { "bbox": [ 106, 419, 505, 430 ], "score": 1.0, "content": "can now output non-deterministic foreground contents even for a fixed input vector which results", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 429, 505, 441 ], "spans": [ { "bbox": [ 106, 429, 505, 441 ], "score": 1.0, "content": "in better visual quality and higher diversity of generated defects. In the main text, the scores of the", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 441, 206, 452 ], "spans": [ { "bbox": [ 105, 441, 206, 452 ], "score": 1.0, "content": "mixture set are reported.", "type": "text" } ], "index": 21 } ], "index": 17 }, { "type": "table", "bbox": [ 106, 481, 512, 600 ], "blocks": [ { "type": "table_caption", "bbox": [ 182, 469, 426, 480 ], "group_id": 1, "lines": [ { "bbox": [ 183, 469, 428, 482 ], "spans": [ { "bbox": [ 183, 469, 428, 482 ], "score": 1.0, "content": "Table 15: Ablation study with regard to FID and KID scores.", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "table_body", "bbox": [ 106, 481, 512, 600 ], "group_id": 1, "lines": [ { "bbox": [ 106, 481, 512, 600 ], "spans": [ { "bbox": [ 106, 481, 512, 600 ], "score": 0.983, "html": "
FID↓KID↓
LatentReferenceMixLatentReferenceMix
(a) Baseline StarGAN v237.7337.9937.700.0130.0130.013
(b)+ Style-Content branches43.9032.6133.360.0170.0110.011
(c)+Foreground classifier37.1432.3427.690.0140.0110.008
(d) + Background classifier34.1232.5030.230.0110.0110.010
(e)+ Separately decoding foreground and background in G48.5238.1134.790.0170.0150.011
(f) + Anchor foreground domain (e.g. No rmal)43.6637.4532.150.0190.0150.011
33.050.0090.011
(g)+ Noise injection in Mapping Network34.4229.730.009
", "type": "table", "image_path": "d91c27f754a46e96b8e6645e98fa2909f79399eb5f16398ddc585747f2f70c05.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 106, 481, 512, 520.6666666666666 ], "spans": [], "index": 23 }, { "bbox": [ 106, 520.6666666666666, 512, 560.3333333333333 ], "spans": [], "index": 24 }, { "bbox": [ 106, 560.3333333333333, 512, 599.9999999999999 ], "spans": [], "index": 25 } ] } ], "index": 23.0 }, { "type": "title", "bbox": [ 107, 623, 441, 635 ], "lines": [ { "bbox": [ 106, 623, 442, 635 ], "spans": [ { "bbox": [ 106, 623, 442, 635 ], "score": 1.0, "content": "E.4 ADDITIONAL RESULTS ON THE MVTEC ANOMALY DETECTION DATASET", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 106, 643, 505, 732 ], "lines": [ { "bbox": [ 106, 644, 505, 656 ], "spans": [ { "bbox": [ 106, 644, 505, 656 ], "score": 1.0, "content": "The MVTec Anomaly Detection dataset (Bergmann et al., 2019) contains 15 different object and", "type": "text" } ], "index": 27 }, { "bbox": [ 104, 653, 505, 668 ], "spans": [ { "bbox": [ 104, 653, 505, 668 ], "score": 1.0, "content": "texture categories for anomaly detection. The dataset is formed of non-defective image for training", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 666, 504, 678 ], "spans": [ { "bbox": [ 106, 666, 504, 678 ], "score": 1.0, "content": "and both non-defective and defective images with various kinds of defects for testing. The pixel-level", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "annotations of all defective images are also provided. It is worth noting that the MVTec Anomaly", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 686, 505, 702 ], "spans": [ { "bbox": [ 105, 686, 505, 702 ], "score": 1.0, "content": "Detection dataset is relatively small scale in number of images, where the number of training images", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "is ranging from 60 to 391. Moreover, the number of defective images for each defect category in the", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 710, 505, 722 ], "spans": [ { "bbox": [ 105, 710, 505, 722 ], "score": 1.0, "content": "test set is varying only from 8 to 30, which is relatively limited considering the sophisticated pattern", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 720, 151, 732 ], "spans": [ { "bbox": [ 105, 720, 151, 732 ], "score": 1.0, "content": "of defects.", "type": "text" } ], "index": 34 } ], "index": 30.5 } ], "page_idx": 19, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 138 ], "lines": [], "index": 2, "bbox_fs": [ 104, 83, 507, 140 ], "lines_deleted": true }, { "type": "table", "bbox": [ 106, 200, 503, 309 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 155, 505, 200 ], "group_id": 0, "lines": [ { "bbox": [ 105, 155, 506, 169 ], "spans": [ { "bbox": [ 105, 155, 506, 169 ], "score": 1.0, "content": "Table 14: Quantitative comparison of DT-GAN with baseline image synthesis methods using FID", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 167, 506, 180 ], "spans": [ { "bbox": [ 105, 167, 506, 180 ], "score": 1.0, "content": "and KID. Note that the reported values are not comparable between columns, because they are", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 177, 506, 190 ], "spans": [ { "bbox": [ 105, 177, 506, 190 ], "score": 1.0, "content": "calculated on different training sets. The scores of StarGAN v2 on single products are omitted", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 188, 488, 201 ], "spans": [ { "bbox": [ 105, 188, 488, 201 ], "score": 1.0, "content": "because generating images with specified background is not possible due to its network design.", "type": "text" } ], "index": 8 } ], "index": 6.5 }, { "type": "table_body", "bbox": [ 106, 200, 503, 309 ], "group_id": 0, "lines": [ { "bbox": [ 106, 200, 503, 309 ], "spans": [ { "bbox": [ 106, 200, 503, 309 ], "score": 0.982, "html": "
MethodFID↓KID↓
ABCAllABCAll
Mokady et al. (2020)68.6966.9036.2158.630.0500.0360.0300.036
StarGAN v2--137.701-10.013
StyleGAN v290.1052.95138.0935.340.0720.0270.1860.013
StyleGAN v2 + ADA149.6642.75135.6976.160.1380.0190.1910.055
BigGAN235.66192.89193.61151.430.2480.1990.2760.115
BigGAN + DiffAug218.74134.41270.89155.880.2200.1210.3780.099
Ours58.4336.4422.6829.730.0250.0130.0120.009
", "type": "table", "image_path": "45a92e9be03f105fca9e42ed49982648d41734c531d7e3f15ce1a34f30d32186.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 106, 200, 503, 236.33333333333334 ], "spans": [], "index": 9 }, { "bbox": [ 106, 236.33333333333334, 503, 272.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 106, 272.6666666666667, 503, 309.0 ], "spans": [], "index": 11 } ] } ], "index": 8.25 }, { "type": "title", "bbox": [ 107, 331, 385, 343 ], "lines": [ { "bbox": [ 105, 330, 386, 345 ], "spans": [ { "bbox": [ 105, 330, 386, 345 ], "score": 1.0, "content": "E.3 ABLATION STUDY WITH REGARD TO FID AND KID SCORES", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 106, 351, 505, 451 ], "lines": [ { "bbox": [ 106, 352, 505, 365 ], "spans": [ { "bbox": [ 106, 352, 505, 365 ], "score": 1.0, "content": "We report the FID and KID scores of the ablation study in Table 15. We notice that both subnetworks", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 361, 506, 378 ], "spans": [ { "bbox": [ 104, 361, 506, 378 ], "score": 1.0, "content": "show positive correlation to each modification except for structural change as in (a) and (e) . Among", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 375, 505, 387 ], "spans": [ { "bbox": [ 106, 375, 505, 387 ], "score": 1.0, "content": "the two subnetworks, the reference-guided subnetwork outperforms the latent-guided one in the", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 385, 505, 397 ], "spans": [ { "bbox": [ 106, 385, 505, 397 ], "score": 1.0, "content": "beginning, which is due to the fact that transferring existing contents is easier than generating them", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 396, 505, 408 ], "spans": [ { "bbox": [ 106, 396, 505, 408 ], "score": 1.0, "content": "from random noise. This effect is also observed in Figure 5. However, the performance of the latent-", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 407, 505, 420 ], "spans": [ { "bbox": [ 106, 407, 505, 420 ], "score": 1.0, "content": "guided subnetwork improves significantly after applying per-pixel noise injection. The subnetwork", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 419, 505, 430 ], "spans": [ { "bbox": [ 106, 419, 505, 430 ], "score": 1.0, "content": "can now output non-deterministic foreground contents even for a fixed input vector which results", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 429, 505, 441 ], "spans": [ { "bbox": [ 106, 429, 505, 441 ], "score": 1.0, "content": "in better visual quality and higher diversity of generated defects. In the main text, the scores of the", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 441, 206, 452 ], "spans": [ { "bbox": [ 105, 441, 206, 452 ], "score": 1.0, "content": "mixture set are reported.", "type": "text" } ], "index": 21 } ], "index": 17, "bbox_fs": [ 104, 352, 506, 452 ] }, { "type": "table", "bbox": [ 106, 481, 512, 600 ], "blocks": [ { "type": "table_caption", "bbox": [ 182, 469, 426, 480 ], "group_id": 1, "lines": [ { "bbox": [ 183, 469, 428, 482 ], "spans": [ { "bbox": [ 183, 469, 428, 482 ], "score": 1.0, "content": "Table 15: Ablation study with regard to FID and KID scores.", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "table_body", "bbox": [ 106, 481, 512, 600 ], "group_id": 1, "lines": [ { "bbox": [ 106, 481, 512, 600 ], "spans": [ { "bbox": [ 106, 481, 512, 600 ], "score": 0.983, "html": "
FID↓KID↓
LatentReferenceMixLatentReferenceMix
(a) Baseline StarGAN v237.7337.9937.700.0130.0130.013
(b)+ Style-Content branches43.9032.6133.360.0170.0110.011
(c)+Foreground classifier37.1432.3427.690.0140.0110.008
(d) + Background classifier34.1232.5030.230.0110.0110.010
(e)+ Separately decoding foreground and background in G48.5238.1134.790.0170.0150.011
(f) + Anchor foreground domain (e.g. No rmal)43.6637.4532.150.0190.0150.011
33.050.0090.011
(g)+ Noise injection in Mapping Network34.4229.730.009
", "type": "table", "image_path": "d91c27f754a46e96b8e6645e98fa2909f79399eb5f16398ddc585747f2f70c05.jpg" } ] } ], "index": 24, "virtual_lines": [ { "bbox": [ 106, 481, 512, 520.6666666666666 ], "spans": [], "index": 23 }, { "bbox": [ 106, 520.6666666666666, 512, 560.3333333333333 ], "spans": [], "index": 24 }, { "bbox": [ 106, 560.3333333333333, 512, 599.9999999999999 ], "spans": [], "index": 25 } ] } ], "index": 23.0 }, { "type": "title", "bbox": [ 107, 623, 441, 635 ], "lines": [ { "bbox": [ 106, 623, 442, 635 ], "spans": [ { "bbox": [ 106, 623, 442, 635 ], "score": 1.0, "content": "E.4 ADDITIONAL RESULTS ON THE MVTEC ANOMALY DETECTION DATASET", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 106, 643, 505, 732 ], "lines": [ { "bbox": [ 106, 644, 505, 656 ], "spans": [ { "bbox": [ 106, 644, 505, 656 ], "score": 1.0, "content": "The MVTec Anomaly Detection dataset (Bergmann et al., 2019) contains 15 different object and", "type": "text" } ], "index": 27 }, { "bbox": [ 104, 653, 505, 668 ], "spans": [ { "bbox": [ 104, 653, 505, 668 ], "score": 1.0, "content": "texture categories for anomaly detection. The dataset is formed of non-defective image for training", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 666, 504, 678 ], "spans": [ { "bbox": [ 106, 666, 504, 678 ], "score": 1.0, "content": "and both non-defective and defective images with various kinds of defects for testing. The pixel-level", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "annotations of all defective images are also provided. It is worth noting that the MVTec Anomaly", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 686, 505, 702 ], "spans": [ { "bbox": [ 105, 686, 505, 702 ], "score": 1.0, "content": "Detection dataset is relatively small scale in number of images, where the number of training images", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 699, 505, 711 ], "spans": [ { "bbox": [ 105, 699, 505, 711 ], "score": 1.0, "content": "is ranging from 60 to 391. Moreover, the number of defective images for each defect category in the", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 710, 505, 722 ], "spans": [ { "bbox": [ 105, 710, 505, 722 ], "score": 1.0, "content": "test set is varying only from 8 to 30, which is relatively limited considering the sophisticated pattern", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 720, 151, 732 ], "spans": [ { "bbox": [ 105, 720, 151, 732 ], "score": 1.0, "content": "of defects.", "type": "text" } ], "index": 34 } ], "index": 30.5, "bbox_fs": [ 104, 644, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 170 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "We conducted image synthesis experiments on a subset of MVTec Anomaly Detection dataset,", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "where we selected four texture categories: Carpet, Leather, Wood and Tile for our targeted sce-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 506, 118 ], "spans": [ { "bbox": [ 105, 104, 506, 118 ], "score": 1.0, "content": "nario i.e. surface defects. Furthermore, we aggregated some of the original defect types defined in", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 505, 128 ], "score": 1.0, "content": "the MVTec Anomaly Detection dataset into scratches and spots according to their visual ap-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 126, 505, 141 ], "spans": [ { "bbox": [ 105, 126, 505, 141 ], "score": 1.0, "content": "pearance. We then simply added the subset of the MVTec Anomaly Detection dataset to the training", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 506, 150 ], "spans": [ { "bbox": [ 105, 137, 506, 150 ], "score": 1.0, "content": "set together with the SDI dataset for training DT-GAN. Details of the resulting dataset are shown in", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 161 ], "spans": [ { "bbox": [ 105, 147, 505, 161 ], "score": 1.0, "content": "Table 17. Note that the small scale of available data posts a major challenge for training generative", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 141, 171 ], "spans": [ { "bbox": [ 105, 159, 141, 171 ], "score": 1.0, "content": "models.", "type": "text" } ], "index": 7 } ], "index": 3.5 }, { "type": "text", "bbox": [ 107, 176, 505, 242 ], "lines": [ { "bbox": [ 105, 176, 505, 188 ], "spans": [ { "bbox": [ 105, 176, 505, 188 ], "score": 1.0, "content": "Quantitative Evaluation. We present additional quantitative results on the subset of the MVTec", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 186, 505, 201 ], "spans": [ { "bbox": [ 105, 186, 505, 201 ], "score": 1.0, "content": "Anomaly Detection dataset in Table 16, following the same evaluation setup as described in Ap-", "type": "text" } ], "index": 9 }, { "bbox": [ 104, 197, 506, 211 ], "spans": [ { "bbox": [ 104, 197, 506, 211 ], "score": 1.0, "content": "pendix C. As shown in Table 16, our method achieves the best scores in Carpet and Wood, which", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 209, 506, 222 ], "spans": [ { "bbox": [ 104, 209, 506, 222 ], "score": 1.0, "content": "supports our claim that DT-GAN generates synthetic images with higher fidelity and more diverse", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 219, 506, 233 ], "spans": [ { "bbox": [ 104, 219, 506, 233 ], "score": 1.0, "content": "defect. However, we also observe that StyleGAN v2 seems to outperform our method in Leather", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 231, 145, 243 ], "spans": [ { "bbox": [ 105, 231, 145, 243 ], "score": 1.0, "content": "and Tile.", "type": "text" } ], "index": 13 } ], "index": 10.5 }, { "type": "text", "bbox": [ 108, 248, 503, 270 ], "lines": [ { "bbox": [ 106, 247, 505, 260 ], "spans": [ { "bbox": [ 106, 247, 505, 260 ], "score": 1.0, "content": "Please note that FID and KID are not optimized to evaluate such a small dataset, there the results", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 259, 358, 271 ], "spans": [ { "bbox": [ 106, 259, 358, 271 ], "score": 1.0, "content": "should only be interpreted together with the qualitative results.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 107, 275, 505, 309 ], "lines": [ { "bbox": [ 105, 274, 505, 289 ], "spans": [ { "bbox": [ 105, 274, 505, 289 ], "score": 1.0, "content": "Note the we again omit the FID and KID of StarGAN v2 because it is not cable of generating images", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 286, 506, 299 ], "spans": [ { "bbox": [ 105, 286, 506, 299 ], "score": 1.0, "content": "for a specified product due to the ‘identity-shift’, which is also explained in detail in the qualitative", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 297, 153, 310 ], "spans": [ { "bbox": [ 105, 297, 153, 310 ], "score": 1.0, "content": "evaluation.", "type": "text" } ], "index": 18 } ], "index": 17 }, { "type": "table", "bbox": [ 119, 359, 489, 449 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 326, 504, 358 ], "group_id": 0, "lines": [ { "bbox": [ 106, 326, 505, 338 ], "spans": [ { "bbox": [ 106, 326, 505, 338 ], "score": 1.0, "content": "Table 16: Quantitative comparison of DT-GAN with baseline image synthesis methods using FID", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 337, 505, 349 ], "spans": [ { "bbox": [ 106, 337, 505, 349 ], "score": 1.0, "content": "and KID. Note that the reported values are not comparable between columns, because they were", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 348, 252, 361 ], "spans": [ { "bbox": [ 105, 348, 252, 361 ], "score": 1.0, "content": "calculated on different training sets.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "table_body", "bbox": [ 119, 359, 489, 449 ], "group_id": 0, "lines": [ { "bbox": [ 119, 359, 489, 449 ], "spans": [ { "bbox": [ 119, 359, 489, 449 ], "score": 0.983, "html": "
MethodFID↓KID↓
CarpetLeatherTileWoodCarpetLeatherTileWood
Mokady (2020)41.8760.26275.1281.710.040.030.290.04
StarGAN v211=1
StyleGAN v251.3751.60225.96140.010.050.030.230.12
BigGAN + DiffAug34.47101.70391.54113.320.030.070.420.07
Ours22.7986.13321.3575.830.010.070.360.03
", "type": "table", "image_path": "48579259fe7630f714f2c1e1b5470350c7a4af04e5a533325a61185f9ce9ab51.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 119, 359, 489, 389.0 ], "spans": [], "index": 22 }, { "bbox": [ 119, 389.0, 489, 419.0 ], "spans": [], "index": 23 }, { "bbox": [ 119, 419.0, 489, 449.0 ], "spans": [], "index": 24 } ] } ], "index": 21.5 }, { "type": "text", "bbox": [ 108, 467, 503, 490 ], "lines": [ { "bbox": [ 107, 467, 504, 480 ], "spans": [ { "bbox": [ 107, 467, 504, 480 ], "score": 1.0, "content": "Qualitative Evaluation. For qualitative results, we again discuss the ‘latent-guided’ and ‘reference-", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 479, 223, 491 ], "spans": [ { "bbox": [ 106, 479, 223, 491 ], "score": 1.0, "content": "guided’ synthesis separately.", "type": "text" } ], "index": 26 } ], "index": 25.5 }, { "type": "text", "bbox": [ 107, 495, 505, 704 ], "lines": [ { "bbox": [ 106, 495, 505, 508 ], "spans": [ { "bbox": [ 106, 495, 505, 508 ], "score": 1.0, "content": "We present the ‘latent-guided’ image synthesis results of StyleGAN v2 in Figure 11 and Figure 12", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 506, 505, 519 ], "spans": [ { "bbox": [ 106, 506, 505, 519 ], "score": 1.0, "content": "and BigGAN in Figure 13 and Figure 14. The results are acquired by training one model for each", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 517, 505, 530 ], "spans": [ { "bbox": [ 105, 517, 505, 530 ], "score": 1.0, "content": "product and then generating 16 images from randomly sampled latent codes from each of them.", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 528, 505, 541 ], "spans": [ { "bbox": [ 105, 528, 505, 541 ], "score": 1.0, "content": "As pointed out in Section 4.1.2, both methods can not adapt well on small dataset. They suffer", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 538, 505, 553 ], "spans": [ { "bbox": [ 105, 538, 505, 553 ], "score": 1.0, "content": "from model collapsing and show signs of overfitting by generating images similar to the training", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 550, 505, 563 ], "spans": [ { "bbox": [ 105, 550, 505, 563 ], "score": 1.0, "content": "data. For example, StyleGAN v2 generates images either with no clear defect or identical to the", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 561, 506, 574 ], "spans": [ { "bbox": [ 105, 561, 506, 574 ], "score": 1.0, "content": "training set (e.g., Leather in Figure 11 and Product B in Figure 12). The overfitting we observe", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 572, 506, 585 ], "spans": [ { "bbox": [ 105, 572, 506, 585 ], "score": 1.0, "content": "here also explains the better FID and KID scores in Table 16. For Tile, we can see clear signs of", "type": "text" } ], "index": 34 }, { "bbox": [ 104, 582, 505, 597 ], "spans": [ { "bbox": [ 104, 582, 505, 597 ], "score": 1.0, "content": "mode collapse in the generated Tile images of StyleGAN v2. Similarly, BigGAN produces images", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 594, 506, 608 ], "spans": [ { "bbox": [ 105, 594, 506, 608 ], "score": 1.0, "content": "with single mode and abnormal patterns (e.g., grid structure and gray edges). Unlike StyleGAN v2", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 604, 505, 618 ], "spans": [ { "bbox": [ 105, 604, 505, 618 ], "score": 1.0, "content": "and BigGAN, StarGAN v2 and our method both require images as input (i.e. Source). Therefore,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 616, 505, 629 ], "spans": [ { "bbox": [ 105, 616, 505, 629 ], "score": 1.0, "content": "we randomly sampled two Normal images and applied eight defects which are generated from", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 627, 505, 640 ], "spans": [ { "bbox": [ 105, 627, 505, 640 ], "score": 1.0, "content": "randomly sampled latent codes to each of them. As seen in Figure 15 and Figure 16, StarGAN v2", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 637, 506, 651 ], "spans": [ { "bbox": [ 104, 637, 506, 651 ], "score": 1.0, "content": "fails to preserve the background from the given input images due to the highly entangled FG and", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 649, 505, 662 ], "spans": [ { "bbox": [ 105, 649, 505, 662 ], "score": 1.0, "content": "BG. Also it fails to generates legit and diverse defects without separately modeling the style and the", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 660, 504, 673 ], "spans": [ { "bbox": [ 105, 660, 504, 673 ], "score": 1.0, "content": "content. In contrast to aforementioned methods, our DT-GAN produces images with higher fidelity", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 671, 505, 684 ], "spans": [ { "bbox": [ 105, 671, 505, 684 ], "score": 1.0, "content": "and more diversity in defect patterns as shown in Figure 17 and Figure 18. We believe this again", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 682, 505, 694 ], "spans": [ { "bbox": [ 105, 682, 505, 694 ], "score": 1.0, "content": "prove the importance of style-content separation and FG/BG disentanglement, which we introduce", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 693, 167, 703 ], "spans": [ { "bbox": [ 105, 693, 167, 703 ], "score": 1.0, "content": "in Section 3.1.", "type": "text" } ], "index": 45 } ], "index": 36 }, { "type": "text", "bbox": [ 106, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 709, 504, 722 ], "spans": [ { "bbox": [ 106, 709, 504, 722 ], "score": 1.0, "content": "For ‘reference-guided’ image synthesis, the results of Mokady et al. (2020) are shown in Figure 19", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 720, 506, 734 ], "spans": [ { "bbox": [ 105, 720, 506, 734 ], "score": 1.0, "content": "and Figure 20 while the results of StarGAN v2 are in Figure 21 and Figure 22. We can observe a", "type": "text" } ], "index": 47 } ], "index": 46.5 } ], "page_idx": 20, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 298, 750, 312, 765 ], "spans": [ { "bbox": [ 298, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 170 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "We conducted image synthesis experiments on a subset of MVTec Anomaly Detection dataset,", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "where we selected four texture categories: Carpet, Leather, Wood and Tile for our targeted sce-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 506, 118 ], "spans": [ { "bbox": [ 105, 104, 506, 118 ], "score": 1.0, "content": "nario i.e. surface defects. Furthermore, we aggregated some of the original defect types defined in", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 505, 128 ], "score": 1.0, "content": "the MVTec Anomaly Detection dataset into scratches and spots according to their visual ap-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 126, 505, 141 ], "spans": [ { "bbox": [ 105, 126, 505, 141 ], "score": 1.0, "content": "pearance. We then simply added the subset of the MVTec Anomaly Detection dataset to the training", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 506, 150 ], "spans": [ { "bbox": [ 105, 137, 506, 150 ], "score": 1.0, "content": "set together with the SDI dataset for training DT-GAN. Details of the resulting dataset are shown in", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 161 ], "spans": [ { "bbox": [ 105, 147, 505, 161 ], "score": 1.0, "content": "Table 17. Note that the small scale of available data posts a major challenge for training generative", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 141, 171 ], "spans": [ { "bbox": [ 105, 159, 141, 171 ], "score": 1.0, "content": "models.", "type": "text" } ], "index": 7 } ], "index": 3.5, "bbox_fs": [ 105, 82, 506, 171 ] }, { "type": "text", "bbox": [ 107, 176, 505, 242 ], "lines": [ { "bbox": [ 105, 176, 505, 188 ], "spans": [ { "bbox": [ 105, 176, 505, 188 ], "score": 1.0, "content": "Quantitative Evaluation. We present additional quantitative results on the subset of the MVTec", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 186, 505, 201 ], "spans": [ { "bbox": [ 105, 186, 505, 201 ], "score": 1.0, "content": "Anomaly Detection dataset in Table 16, following the same evaluation setup as described in Ap-", "type": "text" } ], "index": 9 }, { "bbox": [ 104, 197, 506, 211 ], "spans": [ { "bbox": [ 104, 197, 506, 211 ], "score": 1.0, "content": "pendix C. As shown in Table 16, our method achieves the best scores in Carpet and Wood, which", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 209, 506, 222 ], "spans": [ { "bbox": [ 104, 209, 506, 222 ], "score": 1.0, "content": "supports our claim that DT-GAN generates synthetic images with higher fidelity and more diverse", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 219, 506, 233 ], "spans": [ { "bbox": [ 104, 219, 506, 233 ], "score": 1.0, "content": "defect. However, we also observe that StyleGAN v2 seems to outperform our method in Leather", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 231, 145, 243 ], "spans": [ { "bbox": [ 105, 231, 145, 243 ], "score": 1.0, "content": "and Tile.", "type": "text" } ], "index": 13 } ], "index": 10.5, "bbox_fs": [ 104, 176, 506, 243 ] }, { "type": "text", "bbox": [ 108, 248, 503, 270 ], "lines": [ { "bbox": [ 106, 247, 505, 260 ], "spans": [ { "bbox": [ 106, 247, 505, 260 ], "score": 1.0, "content": "Please note that FID and KID are not optimized to evaluate such a small dataset, there the results", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 259, 358, 271 ], "spans": [ { "bbox": [ 106, 259, 358, 271 ], "score": 1.0, "content": "should only be interpreted together with the qualitative results.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 106, 247, 505, 271 ] }, { "type": "text", "bbox": [ 107, 275, 505, 309 ], "lines": [ { "bbox": [ 105, 274, 505, 289 ], "spans": [ { "bbox": [ 105, 274, 505, 289 ], "score": 1.0, "content": "Note the we again omit the FID and KID of StarGAN v2 because it is not cable of generating images", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 286, 506, 299 ], "spans": [ { "bbox": [ 105, 286, 506, 299 ], "score": 1.0, "content": "for a specified product due to the ‘identity-shift’, which is also explained in detail in the qualitative", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 297, 153, 310 ], "spans": [ { "bbox": [ 105, 297, 153, 310 ], "score": 1.0, "content": "evaluation.", "type": "text" } ], "index": 18 } ], "index": 17, "bbox_fs": [ 105, 274, 506, 310 ] }, { "type": "table", "bbox": [ 119, 359, 489, 449 ], "blocks": [ { "type": "table_caption", "bbox": [ 108, 326, 504, 358 ], "group_id": 0, "lines": [ { "bbox": [ 106, 326, 505, 338 ], "spans": [ { "bbox": [ 106, 326, 505, 338 ], "score": 1.0, "content": "Table 16: Quantitative comparison of DT-GAN with baseline image synthesis methods using FID", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 337, 505, 349 ], "spans": [ { "bbox": [ 106, 337, 505, 349 ], "score": 1.0, "content": "and KID. Note that the reported values are not comparable between columns, because they were", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 348, 252, 361 ], "spans": [ { "bbox": [ 105, 348, 252, 361 ], "score": 1.0, "content": "calculated on different training sets.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "table_body", "bbox": [ 119, 359, 489, 449 ], "group_id": 0, "lines": [ { "bbox": [ 119, 359, 489, 449 ], "spans": [ { "bbox": [ 119, 359, 489, 449 ], "score": 0.983, "html": "
MethodFID↓KID↓
CarpetLeatherTileWoodCarpetLeatherTileWood
Mokady (2020)41.8760.26275.1281.710.040.030.290.04
StarGAN v211=1
StyleGAN v251.3751.60225.96140.010.050.030.230.12
BigGAN + DiffAug34.47101.70391.54113.320.030.070.420.07
Ours22.7986.13321.3575.830.010.070.360.03
", "type": "table", "image_path": "48579259fe7630f714f2c1e1b5470350c7a4af04e5a533325a61185f9ce9ab51.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 119, 359, 489, 389.0 ], "spans": [], "index": 22 }, { "bbox": [ 119, 389.0, 489, 419.0 ], "spans": [], "index": 23 }, { "bbox": [ 119, 419.0, 489, 449.0 ], "spans": [], "index": 24 } ] } ], "index": 21.5 }, { "type": "text", "bbox": [ 108, 467, 503, 490 ], "lines": [ { "bbox": [ 107, 467, 504, 480 ], "spans": [ { "bbox": [ 107, 467, 504, 480 ], "score": 1.0, "content": "Qualitative Evaluation. For qualitative results, we again discuss the ‘latent-guided’ and ‘reference-", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 479, 223, 491 ], "spans": [ { "bbox": [ 106, 479, 223, 491 ], "score": 1.0, "content": "guided’ synthesis separately.", "type": "text" } ], "index": 26 } ], "index": 25.5, "bbox_fs": [ 106, 467, 504, 491 ] }, { "type": "text", "bbox": [ 107, 495, 505, 704 ], "lines": [ { "bbox": [ 106, 495, 505, 508 ], "spans": [ { "bbox": [ 106, 495, 505, 508 ], "score": 1.0, "content": "We present the ‘latent-guided’ image synthesis results of StyleGAN v2 in Figure 11 and Figure 12", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 506, 505, 519 ], "spans": [ { "bbox": [ 106, 506, 505, 519 ], "score": 1.0, "content": "and BigGAN in Figure 13 and Figure 14. The results are acquired by training one model for each", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 517, 505, 530 ], "spans": [ { "bbox": [ 105, 517, 505, 530 ], "score": 1.0, "content": "product and then generating 16 images from randomly sampled latent codes from each of them.", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 528, 505, 541 ], "spans": [ { "bbox": [ 105, 528, 505, 541 ], "score": 1.0, "content": "As pointed out in Section 4.1.2, both methods can not adapt well on small dataset. They suffer", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 538, 505, 553 ], "spans": [ { "bbox": [ 105, 538, 505, 553 ], "score": 1.0, "content": "from model collapsing and show signs of overfitting by generating images similar to the training", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 550, 505, 563 ], "spans": [ { "bbox": [ 105, 550, 505, 563 ], "score": 1.0, "content": "data. For example, StyleGAN v2 generates images either with no clear defect or identical to the", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 561, 506, 574 ], "spans": [ { "bbox": [ 105, 561, 506, 574 ], "score": 1.0, "content": "training set (e.g., Leather in Figure 11 and Product B in Figure 12). The overfitting we observe", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 572, 506, 585 ], "spans": [ { "bbox": [ 105, 572, 506, 585 ], "score": 1.0, "content": "here also explains the better FID and KID scores in Table 16. For Tile, we can see clear signs of", "type": "text" } ], "index": 34 }, { "bbox": [ 104, 582, 505, 597 ], "spans": [ { "bbox": [ 104, 582, 505, 597 ], "score": 1.0, "content": "mode collapse in the generated Tile images of StyleGAN v2. Similarly, BigGAN produces images", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 594, 506, 608 ], "spans": [ { "bbox": [ 105, 594, 506, 608 ], "score": 1.0, "content": "with single mode and abnormal patterns (e.g., grid structure and gray edges). Unlike StyleGAN v2", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 604, 505, 618 ], "spans": [ { "bbox": [ 105, 604, 505, 618 ], "score": 1.0, "content": "and BigGAN, StarGAN v2 and our method both require images as input (i.e. Source). Therefore,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 616, 505, 629 ], "spans": [ { "bbox": [ 105, 616, 505, 629 ], "score": 1.0, "content": "we randomly sampled two Normal images and applied eight defects which are generated from", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 627, 505, 640 ], "spans": [ { "bbox": [ 105, 627, 505, 640 ], "score": 1.0, "content": "randomly sampled latent codes to each of them. As seen in Figure 15 and Figure 16, StarGAN v2", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 637, 506, 651 ], "spans": [ { "bbox": [ 104, 637, 506, 651 ], "score": 1.0, "content": "fails to preserve the background from the given input images due to the highly entangled FG and", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 649, 505, 662 ], "spans": [ { "bbox": [ 105, 649, 505, 662 ], "score": 1.0, "content": "BG. Also it fails to generates legit and diverse defects without separately modeling the style and the", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 660, 504, 673 ], "spans": [ { "bbox": [ 105, 660, 504, 673 ], "score": 1.0, "content": "content. In contrast to aforementioned methods, our DT-GAN produces images with higher fidelity", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 671, 505, 684 ], "spans": [ { "bbox": [ 105, 671, 505, 684 ], "score": 1.0, "content": "and more diversity in defect patterns as shown in Figure 17 and Figure 18. We believe this again", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 682, 505, 694 ], "spans": [ { "bbox": [ 105, 682, 505, 694 ], "score": 1.0, "content": "prove the importance of style-content separation and FG/BG disentanglement, which we introduce", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 693, 167, 703 ], "spans": [ { "bbox": [ 105, 693, 167, 703 ], "score": 1.0, "content": "in Section 3.1.", "type": "text" } ], "index": 45 } ], "index": 36, "bbox_fs": [ 104, 495, 506, 703 ] }, { "type": "text", "bbox": [ 106, 709, 504, 732 ], "lines": [ { "bbox": [ 106, 709, 504, 722 ], "spans": [ { "bbox": [ 106, 709, 504, 722 ], "score": 1.0, "content": "For ‘reference-guided’ image synthesis, the results of Mokady et al. (2020) are shown in Figure 19", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 720, 506, 734 ], "spans": [ { "bbox": [ 105, 720, 506, 734 ], "score": 1.0, "content": "and Figure 20 while the results of StarGAN v2 are in Figure 21 and Figure 22. We can observe a", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "clear shift in color in all the outputs from Mokady et al. (2020). Moreover, Mokady et al. (2020)", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 106, 95, 505, 106 ], "spans": [ { "bbox": [ 106, 95, 505, 106 ], "score": 1.0, "content": "can only transfer content between two domains. In order to perform translation from a non-defective", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 105, 506, 117 ], "spans": [ { "bbox": [ 105, 105, 506, 117 ], "score": 1.0, "content": "sample to a defective one, we trained a model for each type of defect and for each product. This sums", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 505, 128 ], "score": 1.0, "content": "up to be 13 models (Scratches and Spots for 6 categories and Scratches only for Tile). The", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 125, 505, 139 ], "spans": [ { "bbox": [ 105, 125, 505, 139 ], "score": 1.0, "content": "results from the intended use within one background domain can be found on the diagonal and are", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 105, 137, 504, 149 ], "spans": [ { "bbox": [ 105, 137, 504, 149 ], "score": 1.0, "content": "marked in red in both Figure 19 and Figure 20. We still show the images that we feed in images from", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 105, 147, 505, 162 ], "spans": [ { "bbox": [ 105, 147, 505, 162 ], "score": 1.0, "content": "other background domains. As expected, the model then fails to preserve the background of given", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 105, 159, 505, 172 ], "spans": [ { "bbox": [ 105, 159, 505, 172 ], "score": 1.0, "content": "source images and introduce artifacts to the outputs. Similarly, StarGAN v2 does not preserve the", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 105, 171, 505, 183 ], "spans": [ { "bbox": [ 105, 171, 505, 183 ], "score": 1.0, "content": "background from the input images. Without style-content separation and FG/BG disentanglement,", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 181, 505, 194 ], "spans": [ { "bbox": [ 105, 181, 505, 194 ], "score": 1.0, "content": "we observe that StarGAN v2 encodes the background characteristics together with the foreground", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 193, 504, 204 ], "spans": [ { "bbox": [ 105, 193, 504, 204 ], "score": 1.0, "content": "content of the reference images, which results in identity-shits in its output images. Moreover, the", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 204, 504, 216 ], "spans": [ { "bbox": [ 105, 204, 504, 216 ], "score": 1.0, "content": "output images either show no clear defect or contain abnormal patterns which sabotage the fidelity.", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 105, 213, 506, 227 ], "spans": [ { "bbox": [ 105, 213, 506, 227 ], "score": 1.0, "content": "On the contrary, our method can faithfully transfer the foreground content of reference images across", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 105, 225, 505, 238 ], "spans": [ { "bbox": [ 105, 225, 505, 238 ], "score": 1.0, "content": "given background of different products as shown in Figure 23 and Figure 24, which demonstrate", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 105, 236, 506, 249 ], "spans": [ { "bbox": [ 105, 236, 506, 249 ], "score": 1.0, "content": "the effectiveness of the style-content separation and FG/BG disentanglement we introduced in", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 106, 247, 156, 258 ], "spans": [ { "bbox": [ 106, 247, 156, 258 ], "score": 1.0, "content": "Section 3.1.", "type": "text", "cross_page": true } ], "index": 15 } ], "index": 46.5, "bbox_fs": [ 105, 709, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 83, 505, 258 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "clear shift in color in all the outputs from Mokady et al. (2020). Moreover, Mokady et al. (2020)", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 95, 505, 106 ], "spans": [ { "bbox": [ 106, 95, 505, 106 ], "score": 1.0, "content": "can only transfer content between two domains. In order to perform translation from a non-defective", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 506, 117 ], "spans": [ { "bbox": [ 105, 105, 506, 117 ], "score": 1.0, "content": "sample to a defective one, we trained a model for each type of defect and for each product. This sums", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 505, 128 ], "score": 1.0, "content": "up to be 13 models (Scratches and Spots for 6 categories and Scratches only for Tile). The", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 125, 505, 139 ], "spans": [ { "bbox": [ 105, 125, 505, 139 ], "score": 1.0, "content": "results from the intended use within one background domain can be found on the diagonal and are", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 504, 149 ], "spans": [ { "bbox": [ 105, 137, 504, 149 ], "score": 1.0, "content": "marked in red in both Figure 19 and Figure 20. We still show the images that we feed in images from", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 505, 162 ], "spans": [ { "bbox": [ 105, 147, 505, 162 ], "score": 1.0, "content": "other background domains. As expected, the model then fails to preserve the background of given", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 159, 505, 172 ], "spans": [ { "bbox": [ 105, 159, 505, 172 ], "score": 1.0, "content": "source images and introduce artifacts to the outputs. Similarly, StarGAN v2 does not preserve the", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 171, 505, 183 ], "spans": [ { "bbox": [ 105, 171, 505, 183 ], "score": 1.0, "content": "background from the input images. Without style-content separation and FG/BG disentanglement,", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 181, 505, 194 ], "spans": [ { "bbox": [ 105, 181, 505, 194 ], "score": 1.0, "content": "we observe that StarGAN v2 encodes the background characteristics together with the foreground", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 193, 504, 204 ], "spans": [ { "bbox": [ 105, 193, 504, 204 ], "score": 1.0, "content": "content of the reference images, which results in identity-shits in its output images. Moreover, the", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 204, 504, 216 ], "spans": [ { "bbox": [ 105, 204, 504, 216 ], "score": 1.0, "content": "output images either show no clear defect or contain abnormal patterns which sabotage the fidelity.", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 213, 506, 227 ], "spans": [ { "bbox": [ 105, 213, 506, 227 ], "score": 1.0, "content": "On the contrary, our method can faithfully transfer the foreground content of reference images across", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 225, 505, 238 ], "spans": [ { "bbox": [ 105, 225, 505, 238 ], "score": 1.0, "content": "given background of different products as shown in Figure 23 and Figure 24, which demonstrate", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 236, 506, 249 ], "spans": [ { "bbox": [ 105, 236, 506, 249 ], "score": 1.0, "content": "the effectiveness of the style-content separation and FG/BG disentanglement we introduced in", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 247, 156, 258 ], "spans": [ { "bbox": [ 106, 247, 156, 258 ], "score": 1.0, "content": "Section 3.1.", "type": "text" } ], "index": 15 } ], "index": 7.5 }, { "type": "text", "bbox": [ 107, 264, 504, 319 ], "lines": [ { "bbox": [ 104, 263, 506, 276 ], "spans": [ { "bbox": [ 104, 263, 506, 276 ], "score": 1.0, "content": "It is also worth noting that our method can perform cross-domain image synthesis even the desired", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 288 ], "spans": [ { "bbox": [ 105, 274, 505, 288 ], "score": 1.0, "content": "combination is not presented in the training set. We demonstrate this on product Tile, which only", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 285, 506, 299 ], "spans": [ { "bbox": [ 105, 285, 506, 299 ], "score": 1.0, "content": "has images with Scratches but no Spots. As shown in Figure 18 and Figure 24, DT-GAN can", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 296, 506, 310 ], "spans": [ { "bbox": [ 105, 296, 506, 310 ], "score": 1.0, "content": "generated spots one given Tile images. However, this kind of transformation is most useful when", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 307, 391, 320 ], "spans": [ { "bbox": [ 105, 307, 391, 320 ], "score": 1.0, "content": "the desired combination is reasonable for the downstream applications.", "type": "text" } ], "index": 20 } ], "index": 18 }, { "type": "text", "bbox": [ 107, 325, 505, 434 ], "lines": [ { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "Limitation and Future Work. We have demonstrated the feasibility of the proposed DT-GAN by", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 336, 505, 348 ], "spans": [ { "bbox": [ 105, 336, 505, 348 ], "score": 1.0, "content": "incorporating more products from the MVTec Anomaly Detection dataset in our training procedure.", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 346, 505, 359 ], "spans": [ { "bbox": [ 105, 346, 505, 359 ], "score": 1.0, "content": "Intensive experiments have shown that the generated images from DT-GAN yielded better results", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 358, 505, 370 ], "spans": [ { "bbox": [ 105, 358, 505, 370 ], "score": 1.0, "content": "compared to the baseline image synthesis methods. However, we noticed that despite the diverse", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 368, 506, 381 ], "spans": [ { "bbox": [ 105, 368, 506, 381 ], "score": 1.0, "content": "patterns of the generated defects, DT-GAN tends to apply the styles learned from the SDI dataset", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 379, 505, 392 ], "spans": [ { "bbox": [ 105, 379, 505, 392 ], "score": 1.0, "content": "also to the samples from the MVTec Anomaly Detection dataset. For example, we can observe some", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 389, 505, 403 ], "spans": [ { "bbox": [ 105, 389, 505, 403 ], "score": 1.0, "content": "”halo” effects in Leather and Wood in Figure 18 and some of the generated scratches in Figure 17", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 401, 505, 414 ], "spans": [ { "bbox": [ 105, 401, 505, 414 ], "score": 1.0, "content": "and Figure 23 are rather weakly pronounced. We hypothesize this can be counteracted by explicitly", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 412, 505, 424 ], "spans": [ { "bbox": [ 105, 412, 505, 424 ], "score": 1.0, "content": "localizing the defect and enforcing the model to learn conditional relationships between ‘styles’ and", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 424, 386, 435 ], "spans": [ { "bbox": [ 106, 424, 386, 435 ], "score": 1.0, "content": "different backgrounds. We aim to address these issues in future work.", "type": "text" } ], "index": 30 } ], "index": 25.5 } ], "page_idx": 21, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 25, 309, 39 ], "spans": [ { "bbox": [ 106, 25, 309, 39 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "22", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 83, 505, 258 ], "lines": [], "index": 7.5, "bbox_fs": [ 105, 82, 506, 258 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 264, 504, 319 ], "lines": [ { "bbox": [ 104, 263, 506, 276 ], "spans": [ { "bbox": [ 104, 263, 506, 276 ], "score": 1.0, "content": "It is also worth noting that our method can perform cross-domain image synthesis even the desired", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 288 ], "spans": [ { "bbox": [ 105, 274, 505, 288 ], "score": 1.0, "content": "combination is not presented in the training set. We demonstrate this on product Tile, which only", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 285, 506, 299 ], "spans": [ { "bbox": [ 105, 285, 506, 299 ], "score": 1.0, "content": "has images with Scratches but no Spots. As shown in Figure 18 and Figure 24, DT-GAN can", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 296, 506, 310 ], "spans": [ { "bbox": [ 105, 296, 506, 310 ], "score": 1.0, "content": "generated spots one given Tile images. However, this kind of transformation is most useful when", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 307, 391, 320 ], "spans": [ { "bbox": [ 105, 307, 391, 320 ], "score": 1.0, "content": "the desired combination is reasonable for the downstream applications.", "type": "text" } ], "index": 20 } ], "index": 18, "bbox_fs": [ 104, 263, 506, 320 ] }, { "type": "text", "bbox": [ 107, 325, 505, 434 ], "lines": [ { "bbox": [ 105, 324, 505, 338 ], "spans": [ { "bbox": [ 105, 324, 505, 338 ], "score": 1.0, "content": "Limitation and Future Work. We have demonstrated the feasibility of the proposed DT-GAN by", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 336, 505, 348 ], "spans": [ { "bbox": [ 105, 336, 505, 348 ], "score": 1.0, "content": "incorporating more products from the MVTec Anomaly Detection dataset in our training procedure.", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 346, 505, 359 ], "spans": [ { "bbox": [ 105, 346, 505, 359 ], "score": 1.0, "content": "Intensive experiments have shown that the generated images from DT-GAN yielded better results", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 358, 505, 370 ], "spans": [ { "bbox": [ 105, 358, 505, 370 ], "score": 1.0, "content": "compared to the baseline image synthesis methods. However, we noticed that despite the diverse", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 368, 506, 381 ], "spans": [ { "bbox": [ 105, 368, 506, 381 ], "score": 1.0, "content": "patterns of the generated defects, DT-GAN tends to apply the styles learned from the SDI dataset", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 379, 505, 392 ], "spans": [ { "bbox": [ 105, 379, 505, 392 ], "score": 1.0, "content": "also to the samples from the MVTec Anomaly Detection dataset. For example, we can observe some", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 389, 505, 403 ], "spans": [ { "bbox": [ 105, 389, 505, 403 ], "score": 1.0, "content": "”halo” effects in Leather and Wood in Figure 18 and some of the generated scratches in Figure 17", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 401, 505, 414 ], "spans": [ { "bbox": [ 105, 401, 505, 414 ], "score": 1.0, "content": "and Figure 23 are rather weakly pronounced. We hypothesize this can be counteracted by explicitly", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 412, 505, 424 ], "spans": [ { "bbox": [ 105, 412, 505, 424 ], "score": 1.0, "content": "localizing the defect and enforcing the model to learn conditional relationships between ‘styles’ and", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 424, 386, 435 ], "spans": [ { "bbox": [ 106, 424, 386, 435 ], "score": 1.0, "content": "different backgrounds. We aim to address these issues in future work.", "type": "text" } ], "index": 30 } ], "index": 25.5, "bbox_fs": [ 105, 324, 506, 435 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 105, 257, 507, 603 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 210, 506, 255 ], "group_id": 0, "lines": [ { "bbox": [ 105, 210, 506, 223 ], "spans": [ { "bbox": [ 105, 210, 506, 223 ], "score": 1.0, "content": "Table 17: Overview of our formation of the MVTec Anomaly Detection sub-dataset. The first", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 221, 505, 233 ], "spans": [ { "bbox": [ 106, 221, 505, 233 ], "score": 1.0, "content": "column represents the original defect types in the MVTec Anomaly Detection dataset while the first", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 233, 505, 245 ], "spans": [ { "bbox": [ 105, 233, 505, 245 ], "score": 1.0, "content": "row stands for the defect types in our targeted scenario. We list the ID of samples we took from the", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 244, 435, 255 ], "spans": [ { "bbox": [ 105, 244, 435, 255 ], "score": 1.0, "content": "MVTec Anomaly Detection dataset and show the number of samples in row Sum.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 105, 257, 507, 603 ], "group_id": 0, "lines": [ { "bbox": [ 105, 257, 507, 603 ], "spans": [ { "bbox": [ 105, 257, 507, 603 ], "score": 0.979, "html": "
(a) Carpet
ScratchesSpots
Color011,012,014,016, 017000,003,004,007,015,018
Thread000-018
Hole-000 - 016
Sum2423
(b) Leather
ScratchesSpots
Color001,003,005,007,009,011,013,015,018000,002,006,008,010,012,014
Cut000 -018 000 - 006,009 - 016-
Fold Glue000 - 002,005-009,011-015,018
Poke003,009,010,016,017000-017
Sum39
48
(c) Tile
ScratchesSpots
Crack000 - 016二
Sum170
(d) Wood
ScratchesSpots
Color003,005
Scratch001-006,008- 010,013 -016,018- 020000 - 016
Hole=000 - 004,006-009
Combined008001,002,009
Sum1912
", "type": "table", "image_path": "e872d4c36a2361439e0ef7091ef852bb00bbd7ba065cebeee40346b44c21a5cb.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 105, 257, 507, 372.3333333333333 ], "spans": [], "index": 4 }, { "bbox": [ 105, 372.3333333333333, 507, 487.66666666666663 ], "spans": [], "index": 5 }, { "bbox": [ 105, 487.66666666666663, 507, 603.0 ], "spans": [], "index": 6 } ] } ], "index": 3.25 } ], "page_idx": 22, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] }, { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 105, 257, 507, 603 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 210, 506, 255 ], "group_id": 0, "lines": [ { "bbox": [ 105, 210, 506, 223 ], "spans": [ { "bbox": [ 105, 210, 506, 223 ], "score": 1.0, "content": "Table 17: Overview of our formation of the MVTec Anomaly Detection sub-dataset. The first", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 221, 505, 233 ], "spans": [ { "bbox": [ 106, 221, 505, 233 ], "score": 1.0, "content": "column represents the original defect types in the MVTec Anomaly Detection dataset while the first", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 233, 505, 245 ], "spans": [ { "bbox": [ 105, 233, 505, 245 ], "score": 1.0, "content": "row stands for the defect types in our targeted scenario. We list the ID of samples we took from the", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 244, 435, 255 ], "spans": [ { "bbox": [ 105, 244, 435, 255 ], "score": 1.0, "content": "MVTec Anomaly Detection dataset and show the number of samples in row Sum.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 105, 257, 507, 603 ], "group_id": 0, "lines": [ { "bbox": [ 105, 257, 507, 603 ], "spans": [ { "bbox": [ 105, 257, 507, 603 ], "score": 0.979, "html": "
(a) Carpet
ScratchesSpots
Color011,012,014,016, 017000,003,004,007,015,018
Thread000-018
Hole-000 - 016
Sum2423
(b) Leather
ScratchesSpots
Color001,003,005,007,009,011,013,015,018000,002,006,008,010,012,014
Cut000 -018 000 - 006,009 - 016-
Fold Glue000 - 002,005-009,011-015,018
Poke003,009,010,016,017000-017
Sum39
48
(c) Tile
ScratchesSpots
Crack000 - 016二
Sum170
(d) Wood
ScratchesSpots
Color003,005
Scratch001-006,008- 010,013 -016,018- 020000 - 016
Hole=000 - 004,006-009
Combined008001,002,009
Sum1912
", "type": "table", "image_path": "e872d4c36a2361439e0ef7091ef852bb00bbd7ba065cebeee40346b44c21a5cb.jpg" } ] } ], "index": 5, "virtual_lines": [ { "bbox": [ 105, 257, 507, 372.3333333333333 ], "spans": [], "index": 4 }, { "bbox": [ 105, 372.3333333333333, 507, 487.66666666666663 ], "spans": [], "index": 5 }, { "bbox": [ 105, 487.66666666666663, 507, 603.0 ], "spans": [], "index": 6 } ] } ], "index": 3.25 } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 245, 114, 379, 123 ], "lines": [ { "bbox": [ 244, 113, 379, 124 ], "spans": [ { "bbox": [ 244, 113, 379, 124 ], "score": 0.997, "content": "Randomly Sampled Defects (Scratches)", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "image", "bbox": [ 149, 124, 464, 658 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 124, 464, 658 ], "group_id": 0, "lines": [ { "bbox": [ 149, 124, 464, 658 ], "spans": [ { "bbox": [ 149, 124, 464, 658 ], "score": 0.971, "type": "image", "image_path": "567f75bf12cbd5a5119634c75240da3d88366422eb7484cba51a63409149e476.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 149, 124, 464, 302.0 ], "spans": [], "index": 1 }, { "bbox": [ 149, 302.0, 464, 480.0 ], "spans": [], "index": 2 }, { "bbox": [ 149, 480.0, 464, 658.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 107, 667, 504, 700 ], "group_id": 0, "lines": [ { "bbox": [ 106, 667, 505, 679 ], "spans": [ { "bbox": [ 106, 667, 505, 679 ], "score": 1.0, "content": "Figure 11: Latent-guided image synthesis results of StyleGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 677, 504, 690 ], "spans": [ { "bbox": [ 105, 677, 504, 690 ], "score": 1.0, "content": "AD dataset. We train a model for each product and generate 16 Scratches images from randomly", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 689, 195, 701 ], "spans": [ { "bbox": [ 105, 689, 195, 701 ], "score": 1.0, "content": "sampled latent codes.", "type": "text" } ], "index": 6 } ], "index": 5 } ], "index": 3.5 } ], "page_idx": 23, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] }, { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 245, 114, 379, 123 ], "lines": [ { "bbox": [ 244, 113, 379, 124 ], "spans": [ { "bbox": [ 244, 113, 379, 124 ], "score": 0.997, "content": "Randomly Sampled Defects (Scratches)", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "image", "bbox": [ 149, 124, 464, 658 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 124, 464, 658 ], "group_id": 0, "lines": [ { "bbox": [ 149, 124, 464, 658 ], "spans": [ { "bbox": [ 149, 124, 464, 658 ], "score": 0.971, "type": "image", "image_path": "567f75bf12cbd5a5119634c75240da3d88366422eb7484cba51a63409149e476.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 149, 124, 464, 302.0 ], "spans": [], "index": 1 }, { "bbox": [ 149, 302.0, 464, 480.0 ], "spans": [], "index": 2 }, { "bbox": [ 149, 480.0, 464, 658.0 ], "spans": [], "index": 3 } ] }, { "type": "image_caption", "bbox": [ 107, 667, 504, 700 ], "group_id": 0, "lines": [ { "bbox": [ 106, 667, 505, 679 ], "spans": [ { "bbox": [ 106, 667, 505, 679 ], "score": 1.0, "content": "Figure 11: Latent-guided image synthesis results of StyleGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 677, 504, 690 ], "spans": [ { "bbox": [ 105, 677, 504, 690 ], "score": 1.0, "content": "AD dataset. We train a model for each product and generate 16 Scratches images from randomly", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 689, 195, 701 ], "spans": [ { "bbox": [ 105, 689, 195, 701 ], "score": 1.0, "content": "sampled latent codes.", "type": "text" } ], "index": 6 } ], "index": 5 } ], "index": 3.5 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 149, 159, 464, 620 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 159, 464, 620 ], "group_id": 0, "lines": [ { "bbox": [ 149, 159, 464, 620 ], "spans": [ { "bbox": [ 149, 159, 464, 620 ], "score": 0.972, "type": "image", "image_path": "3f55519405e1df2af7deb5998aea2cd1baa9a5f4eca9816aac99ce93d6637aef.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 149, 159, 464, 312.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 149, 312.66666666666663, 464, 466.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 149, 466.33333333333326, 464, 619.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 628, 505, 662 ], "group_id": 0, "lines": [ { "bbox": [ 106, 628, 505, 641 ], "spans": [ { "bbox": [ 106, 628, 505, 641 ], "score": 1.0, "content": "Figure 12: Latent-guided image synthesis results of StyleGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 639, 504, 653 ], "spans": [ { "bbox": [ 106, 639, 504, 653 ], "score": 1.0, "content": "AD dataset. We train a model for each product and generate 16 Spots images from randomly", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 651, 195, 663 ], "spans": [ { "bbox": [ 105, 651, 195, 663 ], "score": 1.0, "content": "sampled latent codes.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ], "page_idx": 24, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "25", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 149, 159, 464, 620 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 159, 464, 620 ], "group_id": 0, "lines": [ { "bbox": [ 149, 159, 464, 620 ], "spans": [ { "bbox": [ 149, 159, 464, 620 ], "score": 0.972, "type": "image", "image_path": "3f55519405e1df2af7deb5998aea2cd1baa9a5f4eca9816aac99ce93d6637aef.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 149, 159, 464, 312.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 149, 312.66666666666663, 464, 466.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 149, 466.33333333333326, 464, 619.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 628, 505, 662 ], "group_id": 0, "lines": [ { "bbox": [ 106, 628, 505, 641 ], "spans": [ { "bbox": [ 106, 628, 505, 641 ], "score": 1.0, "content": "Figure 12: Latent-guided image synthesis results of StyleGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 639, 504, 653 ], "spans": [ { "bbox": [ 106, 639, 504, 653 ], "score": 1.0, "content": "AD dataset. We train a model for each product and generate 16 Spots images from randomly", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 651, 195, 663 ], "spans": [ { "bbox": [ 105, 651, 195, 663 ], "score": 1.0, "content": "sampled latent codes.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 149, 123, 464, 660 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 123, 464, 660 ], "group_id": 0, "lines": [ { "bbox": [ 149, 123, 464, 660 ], "spans": [ { "bbox": [ 149, 123, 464, 660 ], "score": 0.963, "type": "image", "image_path": "9af2881df4cd89bcd7f83e40ecac133c4486b383aca6a5f80167e218d4861e0f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 149, 123, 464, 302.0 ], "spans": [], "index": 0 }, { "bbox": [ 149, 302.0, 464, 481.0 ], "spans": [], "index": 1 }, { "bbox": [ 149, 481.0, 464, 660.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 668, 505, 702 ], "group_id": 0, "lines": [ { "bbox": [ 106, 668, 505, 681 ], "spans": [ { "bbox": [ 106, 668, 505, 681 ], "score": 1.0, "content": "Figure 13: Latent-guided image synthesis results of BigGAN with DiffAug on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 679, 505, 693 ], "spans": [ { "bbox": [ 105, 679, 505, 693 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and generate 16 Scratches images", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 690, 257, 703 ], "spans": [ { "bbox": [ 106, 690, 257, 703 ], "score": 1.0, "content": "from randomly sampled latent codes.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ], "page_idx": 25, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 749, 313, 764 ], "spans": [ { "bbox": [ 298, 749, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 149, 123, 464, 660 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 123, 464, 660 ], "group_id": 0, "lines": [ { "bbox": [ 149, 123, 464, 660 ], "spans": [ { "bbox": [ 149, 123, 464, 660 ], "score": 0.963, "type": "image", "image_path": "9af2881df4cd89bcd7f83e40ecac133c4486b383aca6a5f80167e218d4861e0f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 149, 123, 464, 302.0 ], "spans": [], "index": 0 }, { "bbox": [ 149, 302.0, 464, 481.0 ], "spans": [], "index": 1 }, { "bbox": [ 149, 481.0, 464, 660.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 668, 505, 702 ], "group_id": 0, "lines": [ { "bbox": [ 106, 668, 505, 681 ], "spans": [ { "bbox": [ 106, 668, 505, 681 ], "score": 1.0, "content": "Figure 13: Latent-guided image synthesis results of BigGAN with DiffAug on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 679, 505, 693 ], "spans": [ { "bbox": [ 105, 679, 505, 693 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and generate 16 Scratches images", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 690, 257, 703 ], "spans": [ { "bbox": [ 106, 690, 257, 703 ], "score": 1.0, "content": "from randomly sampled latent codes.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 149, 159, 464, 620 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 159, 464, 620 ], "group_id": 0, "lines": [ { "bbox": [ 149, 159, 464, 620 ], "spans": [ { "bbox": [ 149, 159, 464, 620 ], "score": 0.968, "type": "image", "image_path": "6502d31fa4830920177c7a1201405a5d38fd61293a22fa5cc40cc805b4888a23.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 149, 159, 464, 312.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 149, 312.66666666666663, 464, 466.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 149, 466.33333333333326, 464, 619.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 629, 505, 663 ], "group_id": 0, "lines": [ { "bbox": [ 106, 628, 505, 642 ], "spans": [ { "bbox": [ 106, 628, 505, 642 ], "score": 1.0, "content": "Figure 14: Latent-guided image synthesis results of BigGAN with DiffAug on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 639, 505, 653 ], "spans": [ { "bbox": [ 105, 639, 505, 653 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and generate 16 Spots images from", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 651, 235, 664 ], "spans": [ { "bbox": [ 106, 651, 235, 664 ], "score": 1.0, "content": "randomly sampled latent codes.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ], "page_idx": 26, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "27", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 149, 159, 464, 620 ], "blocks": [ { "type": "image_body", "bbox": [ 149, 159, 464, 620 ], "group_id": 0, "lines": [ { "bbox": [ 149, 159, 464, 620 ], "spans": [ { "bbox": [ 149, 159, 464, 620 ], "score": 0.968, "type": "image", "image_path": "6502d31fa4830920177c7a1201405a5d38fd61293a22fa5cc40cc805b4888a23.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 149, 159, 464, 312.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 149, 312.66666666666663, 464, 466.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 149, 466.33333333333326, 464, 619.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 629, 505, 663 ], "group_id": 0, "lines": [ { "bbox": [ 106, 628, 505, 642 ], "spans": [ { "bbox": [ 106, 628, 505, 642 ], "score": 1.0, "content": "Figure 14: Latent-guided image synthesis results of BigGAN with DiffAug on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 639, 505, 653 ], "spans": [ { "bbox": [ 105, 639, 505, 653 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and generate 16 Spots images from", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 651, 235, 664 ], "spans": [ { "bbox": [ 106, 651, 235, 664 ], "score": 1.0, "content": "randomly sampled latent codes.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 129, 113, 484, 645 ], "blocks": [ { "type": "image_body", "bbox": [ 129, 113, 484, 645 ], "group_id": 0, "lines": [ { "bbox": [ 129, 113, 484, 645 ], "spans": [ { "bbox": [ 129, 113, 484, 645 ], "score": 0.964, "type": "image", "image_path": "753dedb2a88181dd1a02122d6d4012b9e01f9fd382ccd9580c8645207ed06d8d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 129, 113, 484, 290.33333333333337 ], "spans": [], "index": 0 }, { "bbox": [ 129, 290.33333333333337, 484, 467.66666666666674 ], "spans": [], "index": 1 }, { "bbox": [ 129, 467.66666666666674, 484, 645.0000000000001 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 653, 506, 709 ], "group_id": 0, "lines": [ { "bbox": [ 106, 654, 505, 664 ], "spans": [ { "bbox": [ 106, 654, 505, 664 ], "score": 1.0, "content": "Figure 15: Latent-guided image synthesis results of StarGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 664, 505, 676 ], "spans": [ { "bbox": [ 106, 664, 505, 676 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 676, 504, 687 ], "spans": [ { "bbox": [ 106, 676, 504, 687 ], "score": 1.0, "content": "from Normal to Scratches. Note that without the style-content separation and the FG/BG", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 685, 505, 699 ], "spans": [ { "bbox": [ 105, 685, 505, 699 ], "score": 1.0, "content": "disentanglement, StarGAN v2 not only fails to preserve the background from the given Source", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 697, 286, 711 ], "spans": [ { "bbox": [ 106, 697, 286, 711 ], "score": 1.0, "content": "image but also fail to generates legit defects.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ], "page_idx": 27, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 312, 763 ], "spans": [ { "bbox": [ 298, 750, 312, 763 ], "score": 1.0, "content": "28", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 129, 113, 484, 645 ], "blocks": [ { "type": "image_body", "bbox": [ 129, 113, 484, 645 ], "group_id": 0, "lines": [ { "bbox": [ 129, 113, 484, 645 ], "spans": [ { "bbox": [ 129, 113, 484, 645 ], "score": 0.964, "type": "image", "image_path": "753dedb2a88181dd1a02122d6d4012b9e01f9fd382ccd9580c8645207ed06d8d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 129, 113, 484, 290.33333333333337 ], "spans": [], "index": 0 }, { "bbox": [ 129, 290.33333333333337, 484, 467.66666666666674 ], "spans": [], "index": 1 }, { "bbox": [ 129, 467.66666666666674, 484, 645.0000000000001 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 653, 506, 709 ], "group_id": 0, "lines": [ { "bbox": [ 106, 654, 505, 664 ], "spans": [ { "bbox": [ 106, 654, 505, 664 ], "score": 1.0, "content": "Figure 15: Latent-guided image synthesis results of StarGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 664, 505, 676 ], "spans": [ { "bbox": [ 106, 664, 505, 676 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 676, 504, 687 ], "spans": [ { "bbox": [ 106, 676, 504, 687 ], "score": 1.0, "content": "from Normal to Scratches. Note that without the style-content separation and the FG/BG", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 685, 505, 699 ], "spans": [ { "bbox": [ 105, 685, 505, 699 ], "score": 1.0, "content": "disentanglement, StarGAN v2 not only fails to preserve the background from the given Source", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 697, 286, 711 ], "spans": [ { "bbox": [ 106, 697, 286, 711 ], "score": 1.0, "content": "image but also fail to generates legit defects.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 130, 111, 484, 647 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 111, 484, 647 ], "group_id": 0, "lines": [ { "bbox": [ 130, 111, 484, 647 ], "spans": [ { "bbox": [ 130, 111, 484, 647 ], "score": 0.966, "type": "image", "image_path": "49e2bc6b1e4d4de05a60e2ef2fb4992d1dc00bdd8bcc62cfa38819c364a1a72e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 130, 111, 484, 289.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 130, 289.66666666666663, 484, 468.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 130, 468.33333333333326, 484, 646.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 654, 505, 709 ], "group_id": 0, "lines": [ { "bbox": [ 106, 655, 505, 667 ], "spans": [ { "bbox": [ 106, 655, 505, 667 ], "score": 1.0, "content": "Figure 16: Latent-guided image synthesis results of StarGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 664, 505, 677 ], "spans": [ { "bbox": [ 106, 664, 505, 677 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 676, 505, 689 ], "spans": [ { "bbox": [ 105, 676, 505, 689 ], "score": 1.0, "content": "from Normal to Spots. Note that without the style-content separation and the FG/BG disentan-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 687, 506, 700 ], "spans": [ { "bbox": [ 106, 687, 506, 700 ], "score": 1.0, "content": "glement, StarGAN v2 not only fails to preserve the background from the given Source image but", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 698, 244, 710 ], "spans": [ { "bbox": [ 106, 698, 244, 710 ], "score": 1.0, "content": "also fail to generates legit defects.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ], "page_idx": 28, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 308, 37 ], "lines": [ { "bbox": [ 107, 26, 308, 38 ], "spans": [ { "bbox": [ 107, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 130, 111, 484, 647 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 111, 484, 647 ], "group_id": 0, "lines": [ { "bbox": [ 130, 111, 484, 647 ], "spans": [ { "bbox": [ 130, 111, 484, 647 ], "score": 0.966, "type": "image", "image_path": "49e2bc6b1e4d4de05a60e2ef2fb4992d1dc00bdd8bcc62cfa38819c364a1a72e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 130, 111, 484, 289.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 130, 289.66666666666663, 484, 468.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 130, 468.33333333333326, 484, 646.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 654, 505, 709 ], "group_id": 0, "lines": [ { "bbox": [ 106, 655, 505, 667 ], "spans": [ { "bbox": [ 106, 655, 505, 667 ], "score": 1.0, "content": "Figure 16: Latent-guided image synthesis results of StarGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 664, 505, 677 ], "spans": [ { "bbox": [ 106, 664, 505, 677 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 676, 505, 689 ], "spans": [ { "bbox": [ 105, 676, 505, 689 ], "score": 1.0, "content": "from Normal to Spots. Note that without the style-content separation and the FG/BG disentan-", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 687, 506, 700 ], "spans": [ { "bbox": [ 106, 687, 506, 700 ], "score": 1.0, "content": "glement, StarGAN v2 not only fails to preserve the background from the given Source image but", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 698, 244, 710 ], "spans": [ { "bbox": [ 106, 698, 244, 710 ], "score": 1.0, "content": "also fail to generates legit defects.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 130, 113, 484, 646 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 113, 484, 646 ], "group_id": 0, "lines": [ { "bbox": [ 130, 113, 484, 646 ], "spans": [ { "bbox": [ 130, 113, 484, 646 ], "score": 0.971, "type": "image", "image_path": "d7665bc9189f97191cb363ff43243d45d9c0eed07ddb8d132a1899df30a87129.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 130, 113, 484, 290.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 130, 290.66666666666663, 484, 468.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 130, 468.33333333333326, 484, 645.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 653, 506, 708 ], "group_id": 0, "lines": [ { "bbox": [ 106, 653, 505, 665 ], "spans": [ { "bbox": [ 106, 653, 505, 665 ], "score": 1.0, "content": "Figure 17: Latent-guided image synthesis results of DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 663, 505, 676 ], "spans": [ { "bbox": [ 105, 663, 505, 676 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 675, 505, 687 ], "spans": [ { "bbox": [ 106, 675, 505, 687 ], "score": 1.0, "content": "from Normal to Scratches. Note the our model takes input Source images as background and", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 686, 505, 699 ], "spans": [ { "bbox": [ 105, 686, 505, 699 ], "score": 1.0, "content": "only synthesizes the foreground defects from randomly sampled latent code compared to StyleGAN", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 696, 177, 709 ], "spans": [ { "bbox": [ 106, 696, 177, 709 ], "score": 1.0, "content": "v2 and BigGAN.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ], "page_idx": 29, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 309, 38 ], "spans": [ { "bbox": [ 106, 26, 309, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 763 ], "spans": [ { "bbox": [ 298, 750, 313, 763 ], "score": 1.0, "content": "30", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 130, 113, 484, 646 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 113, 484, 646 ], "group_id": 0, "lines": [ { "bbox": [ 130, 113, 484, 646 ], "spans": [ { "bbox": [ 130, 113, 484, 646 ], "score": 0.971, "type": "image", "image_path": "d7665bc9189f97191cb363ff43243d45d9c0eed07ddb8d132a1899df30a87129.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 130, 113, 484, 290.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 130, 290.66666666666663, 484, 468.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 130, 468.33333333333326, 484, 645.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 653, 506, 708 ], "group_id": 0, "lines": [ { "bbox": [ 106, 653, 505, 665 ], "spans": [ { "bbox": [ 106, 653, 505, 665 ], "score": 1.0, "content": "Figure 17: Latent-guided image synthesis results of DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 663, 505, 676 ], "spans": [ { "bbox": [ 105, 663, 505, 676 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 675, 505, 687 ], "spans": [ { "bbox": [ 106, 675, 505, 687 ], "score": 1.0, "content": "from Normal to Scratches. Note the our model takes input Source images as background and", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 686, 505, 699 ], "spans": [ { "bbox": [ 105, 686, 505, 699 ], "score": 1.0, "content": "only synthesizes the foreground defects from randomly sampled latent code compared to StyleGAN", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 696, 177, 709 ], "spans": [ { "bbox": [ 106, 696, 177, 709 ], "score": 1.0, "content": "v2 and BigGAN.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 130, 111, 484, 647 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 111, 484, 647 ], "group_id": 0, "lines": [ { "bbox": [ 130, 111, 484, 647 ], "spans": [ { "bbox": [ 130, 111, 484, 647 ], "score": 0.969, "type": "image", "image_path": "6dc33723b61a8a29ea99523303f759e2ec80cca613f5db714539c271f27d80a8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 130, 111, 484, 289.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 130, 289.66666666666663, 484, 468.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 130, 468.33333333333326, 484, 646.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 653, 506, 709 ], "group_id": 0, "lines": [ { "bbox": [ 106, 654, 505, 666 ], "spans": [ { "bbox": [ 106, 654, 505, 666 ], "score": 1.0, "content": "Figure 18: Latent-guided image synthesis results of DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 665, 505, 677 ], "spans": [ { "bbox": [ 106, 665, 505, 677 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 675, 505, 689 ], "spans": [ { "bbox": [ 105, 675, 505, 689 ], "score": 1.0, "content": "from Normal to Spots. Note the our model takes input Source images as background and only", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 686, 505, 700 ], "spans": [ { "bbox": [ 104, 686, 505, 700 ], "score": 1.0, "content": "synthesizes the foreground defects from randomly sampled latent code compared to StyleGAN v2", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 697, 164, 711 ], "spans": [ { "bbox": [ 105, 697, 164, 711 ], "score": 1.0, "content": "and BigGAN.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ], "page_idx": 30, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 309, 38 ], "spans": [ { "bbox": [ 106, 26, 309, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 130, 111, 484, 647 ], "blocks": [ { "type": "image_body", "bbox": [ 130, 111, 484, 647 ], "group_id": 0, "lines": [ { "bbox": [ 130, 111, 484, 647 ], "spans": [ { "bbox": [ 130, 111, 484, 647 ], "score": 0.969, "type": "image", "image_path": "6dc33723b61a8a29ea99523303f759e2ec80cca613f5db714539c271f27d80a8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 130, 111, 484, 289.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 130, 289.66666666666663, 484, 468.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 130, 468.33333333333326, 484, 646.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 653, 506, 709 ], "group_id": 0, "lines": [ { "bbox": [ 106, 654, 505, 666 ], "spans": [ { "bbox": [ 106, 654, 505, 666 ], "score": 1.0, "content": "Figure 18: Latent-guided image synthesis results of DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 665, 505, 677 ], "spans": [ { "bbox": [ 106, 665, 505, 677 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 675, 505, 689 ], "spans": [ { "bbox": [ 105, 675, 505, 689 ], "score": 1.0, "content": "from Normal to Spots. Note the our model takes input Source images as background and only", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 686, 505, 700 ], "spans": [ { "bbox": [ 104, 686, 505, 700 ], "score": 1.0, "content": "synthesizes the foreground defects from randomly sampled latent code compared to StyleGAN v2", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 697, 164, 711 ], "spans": [ { "bbox": [ 105, 697, 164, 711 ], "score": 1.0, "content": "and BigGAN.", "type": "text" } ], "index": 7 } ], "index": 5 } ], "index": 3.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 209, 483, 555 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 209, 483, 555 ], "group_id": 0, "lines": [ { "bbox": [ 132, 209, 483, 555 ], "spans": [ { "bbox": [ 132, 209, 483, 555 ], "score": 0.975, "type": "image", "image_path": "6fbc28db81f825e8f35749ff8c9692af1a0d34358c15a1ea8e335765a76b1ab6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 209, 483, 324.3333333333333 ], "spans": [], "index": 0 }, { "bbox": [ 132, 324.3333333333333, 483, 439.66666666666663 ], "spans": [], "index": 1 }, { "bbox": [ 132, 439.66666666666663, 483, 555.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 106, 567, 505, 579 ], "spans": [ { "bbox": [ 106, 567, 505, 579 ], "score": 1.0, "content": "Figure 19: Reference-guided image synthesis results of Mokady et al. (2020) on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 578, 505, 590 ], "spans": [ { "bbox": [ 106, 578, 505, 590 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and each defect type. Then we translate", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 589, 506, 602 ], "spans": [ { "bbox": [ 105, 589, 506, 602 ], "score": 1.0, "content": "Normal images to Scratches by taking the Source as background and applying the foreground", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 600, 221, 612 ], "spans": [ { "bbox": [ 106, 600, 221, 612 ], "score": 1.0, "content": "defect from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ], "page_idx": 31, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 765 ], "spans": [ { "bbox": [ 298, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 209, 483, 555 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 209, 483, 555 ], "group_id": 0, "lines": [ { "bbox": [ 132, 209, 483, 555 ], "spans": [ { "bbox": [ 132, 209, 483, 555 ], "score": 0.975, "type": "image", "image_path": "6fbc28db81f825e8f35749ff8c9692af1a0d34358c15a1ea8e335765a76b1ab6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 209, 483, 324.3333333333333 ], "spans": [], "index": 0 }, { "bbox": [ 132, 324.3333333333333, 483, 439.66666666666663 ], "spans": [], "index": 1 }, { "bbox": [ 132, 439.66666666666663, 483, 555.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 106, 567, 505, 579 ], "spans": [ { "bbox": [ 106, 567, 505, 579 ], "score": 1.0, "content": "Figure 19: Reference-guided image synthesis results of Mokady et al. (2020) on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 578, 505, 590 ], "spans": [ { "bbox": [ 106, 578, 505, 590 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and each defect type. Then we translate", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 589, 506, 602 ], "spans": [ { "bbox": [ 105, 589, 506, 602 ], "score": 1.0, "content": "Normal images to Scratches by taking the Source as background and applying the foreground", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 600, 221, 612 ], "spans": [ { "bbox": [ 106, 600, 221, 612 ], "score": 1.0, "content": "defect from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 209, 484, 554 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 209, 484, 554 ], "group_id": 0, "lines": [ { "bbox": [ 132, 209, 484, 554 ], "spans": [ { "bbox": [ 132, 209, 484, 554 ], "score": 0.974, "type": "image", "image_path": "52103c7943c67c62f890f5864b5e34df9520df52349661b703bc363d3ec05466.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 209, 484, 324.0 ], "spans": [], "index": 0 }, { "bbox": [ 132, 324.0, 484, 439.0 ], "spans": [], "index": 1 }, { "bbox": [ 132, 439.0, 484, 554.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 106, 567, 505, 579 ], "spans": [ { "bbox": [ 106, 567, 505, 579 ], "score": 1.0, "content": "Figure 20: Reference-guided image synthesis results of Mokady et al. (2020) on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 578, 505, 590 ], "spans": [ { "bbox": [ 106, 578, 505, 590 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and each defect type. Then we translate", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 590, 506, 603 ], "spans": [ { "bbox": [ 105, 590, 506, 603 ], "score": 1.0, "content": "Normal images to Spots by taking the Source as background and applying the foreground defect", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 600, 194, 612 ], "spans": [ { "bbox": [ 105, 600, 194, 612 ], "score": 1.0, "content": "from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ], "page_idx": 32, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 312, 763 ], "spans": [ { "bbox": [ 298, 750, 312, 763 ], "score": 1.0, "content": "33", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 209, 484, 554 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 209, 484, 554 ], "group_id": 0, "lines": [ { "bbox": [ 132, 209, 484, 554 ], "spans": [ { "bbox": [ 132, 209, 484, 554 ], "score": 0.974, "type": "image", "image_path": "52103c7943c67c62f890f5864b5e34df9520df52349661b703bc363d3ec05466.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 209, 484, 324.0 ], "spans": [], "index": 0 }, { "bbox": [ 132, 324.0, 484, 439.0 ], "spans": [], "index": 1 }, { "bbox": [ 132, 439.0, 484, 554.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 106, 567, 505, 579 ], "spans": [ { "bbox": [ 106, 567, 505, 579 ], "score": 1.0, "content": "Figure 20: Reference-guided image synthesis results of Mokady et al. (2020) on the SDI dataset and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 578, 505, 590 ], "spans": [ { "bbox": [ 106, 578, 505, 590 ], "score": 1.0, "content": "the MVTec AD dataset. We train a model for each product and each defect type. Then we translate", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 590, 506, 603 ], "spans": [ { "bbox": [ 105, 590, 506, 603 ], "score": 1.0, "content": "Normal images to Spots by taking the Source as background and applying the foreground defect", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 600, 194, 612 ], "spans": [ { "bbox": [ 105, 600, 194, 612 ], "score": 1.0, "content": "from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 195, 483, 543 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 195, 483, 543 ], "group_id": 0, "lines": [ { "bbox": [ 132, 195, 483, 543 ], "spans": [ { "bbox": [ 132, 195, 483, 543 ], "score": 0.975, "type": "image", "image_path": "cb73ee815eca7466279848e34aba2a7625dd3b27655d8dcff88d57a806fdbab6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 195, 483, 311.0 ], "spans": [], "index": 0 }, { "bbox": [ 132, 311.0, 483, 427.0 ], "spans": [], "index": 1 }, { "bbox": [ 132, 427.0, 483, 543.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 555, 506, 622 ], "group_id": 0, "lines": [ { "bbox": [ 105, 555, 506, 569 ], "spans": [ { "bbox": [ 105, 555, 506, 569 ], "score": 1.0, "content": "Figure 21: Reference-guided image synthesis results StarGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 566, 505, 578 ], "spans": [ { "bbox": [ 106, 566, 505, 578 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 577, 505, 589 ], "spans": [ { "bbox": [ 105, 577, 505, 589 ], "score": 1.0, "content": "from Normal to Scratches by taking the Source as background and applying the foreground", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 587, 505, 601 ], "spans": [ { "bbox": [ 105, 587, 505, 601 ], "score": 1.0, "content": "defect from Reference to it. Note that without the style-content separation and the FG/BG disen-", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 600, 505, 612 ], "spans": [ { "bbox": [ 106, 600, 505, 612 ], "score": 1.0, "content": "tanglement, StarGAN v2 not only fails to preserve the background from the given Source image but", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 610, 244, 623 ], "spans": [ { "bbox": [ 105, 610, 244, 623 ], "score": 1.0, "content": "also fail to generates legit defects.", "type": "text" } ], "index": 8 } ], "index": 5.5 } ], "index": 3.25 } ], "page_idx": 33, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 749, 313, 764 ], "spans": [ { "bbox": [ 298, 749, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 195, 483, 543 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 195, 483, 543 ], "group_id": 0, "lines": [ { "bbox": [ 132, 195, 483, 543 ], "spans": [ { "bbox": [ 132, 195, 483, 543 ], "score": 0.975, "type": "image", "image_path": "cb73ee815eca7466279848e34aba2a7625dd3b27655d8dcff88d57a806fdbab6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 195, 483, 311.0 ], "spans": [], "index": 0 }, { "bbox": [ 132, 311.0, 483, 427.0 ], "spans": [], "index": 1 }, { "bbox": [ 132, 427.0, 483, 543.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 555, 506, 622 ], "group_id": 0, "lines": [ { "bbox": [ 105, 555, 506, 569 ], "spans": [ { "bbox": [ 105, 555, 506, 569 ], "score": 1.0, "content": "Figure 21: Reference-guided image synthesis results StarGAN v2 on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 566, 505, 578 ], "spans": [ { "bbox": [ 106, 566, 505, 578 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 577, 505, 589 ], "spans": [ { "bbox": [ 105, 577, 505, 589 ], "score": 1.0, "content": "from Normal to Scratches by taking the Source as background and applying the foreground", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 587, 505, 601 ], "spans": [ { "bbox": [ 105, 587, 505, 601 ], "score": 1.0, "content": "defect from Reference to it. Note that without the style-content separation and the FG/BG disen-", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 600, 505, 612 ], "spans": [ { "bbox": [ 106, 600, 505, 612 ], "score": 1.0, "content": "tanglement, StarGAN v2 not only fails to preserve the background from the given Source image but", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 610, 244, 623 ], "spans": [ { "bbox": [ 105, 610, 244, 623 ], "score": 1.0, "content": "also fail to generates legit defects.", "type": "text" } ], "index": 8 } ], "index": 5.5 } ], "index": 3.25 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 196, 484, 543 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 196, 484, 543 ], "group_id": 0, "lines": [ { "bbox": [ 132, 196, 484, 543 ], "spans": [ { "bbox": [ 132, 196, 484, 543 ], "score": 0.974, "type": "image", "image_path": "c0757dd5081906d67150a0b1e7071d90d4070ef9a33fedd3004d5d77e8301ddd.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 196, 484, 311.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 132, 311.6666666666667, 484, 427.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 132, 427.33333333333337, 484, 543.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 555, 505, 622 ], "group_id": 0, "lines": [ { "bbox": [ 106, 555, 505, 567 ], "spans": [ { "bbox": [ 106, 555, 505, 567 ], "score": 1.0, "content": "Figure 22: Reference-guided image synthesis results of StarGAN v2 on the SDI dataset and the", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 565, 505, 579 ], "spans": [ { "bbox": [ 105, 565, 505, 579 ], "score": 1.0, "content": "MVTec AD dataset. The model is trained on a joint set of aforementioned datasets and performs", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 576, 504, 590 ], "spans": [ { "bbox": [ 105, 576, 504, 590 ], "score": 1.0, "content": "translation from Normal to Spots by taking the Source as background and applying the fore-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 588, 505, 600 ], "spans": [ { "bbox": [ 105, 588, 505, 600 ], "score": 1.0, "content": "ground defect from Reference to it. Note that without the style-content separation and the FG/BG", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 598, 505, 612 ], "spans": [ { "bbox": [ 106, 598, 505, 612 ], "score": 1.0, "content": "disentanglement, StarGAN v2 not only fails to preserve the background from the given Source", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 610, 286, 622 ], "spans": [ { "bbox": [ 106, 610, 286, 622 ], "score": 1.0, "content": "image but also fail to generates legit defects.", "type": "text" } ], "index": 8 } ], "index": 5.5 } ], "index": 3.25 } ], "page_idx": 34, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 749, 313, 764 ], "spans": [ { "bbox": [ 298, 749, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 196, 484, 543 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 196, 484, 543 ], "group_id": 0, "lines": [ { "bbox": [ 132, 196, 484, 543 ], "spans": [ { "bbox": [ 132, 196, 484, 543 ], "score": 0.974, "type": "image", "image_path": "c0757dd5081906d67150a0b1e7071d90d4070ef9a33fedd3004d5d77e8301ddd.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 196, 484, 311.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 132, 311.6666666666667, 484, 427.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 132, 427.33333333333337, 484, 543.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 555, 505, 622 ], "group_id": 0, "lines": [ { "bbox": [ 106, 555, 505, 567 ], "spans": [ { "bbox": [ 106, 555, 505, 567 ], "score": 1.0, "content": "Figure 22: Reference-guided image synthesis results of StarGAN v2 on the SDI dataset and the", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 565, 505, 579 ], "spans": [ { "bbox": [ 105, 565, 505, 579 ], "score": 1.0, "content": "MVTec AD dataset. The model is trained on a joint set of aforementioned datasets and performs", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 576, 504, 590 ], "spans": [ { "bbox": [ 105, 576, 504, 590 ], "score": 1.0, "content": "translation from Normal to Spots by taking the Source as background and applying the fore-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 588, 505, 600 ], "spans": [ { "bbox": [ 105, 588, 505, 600 ], "score": 1.0, "content": "ground defect from Reference to it. Note that without the style-content separation and the FG/BG", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 598, 505, 612 ], "spans": [ { "bbox": [ 106, 598, 505, 612 ], "score": 1.0, "content": "disentanglement, StarGAN v2 not only fails to preserve the background from the given Source", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 610, 286, 622 ], "spans": [ { "bbox": [ 106, 610, 286, 622 ], "score": 1.0, "content": "image but also fail to generates legit defects.", "type": "text" } ], "index": 8 } ], "index": 5.5 } ], "index": 3.25 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 206, 484, 555 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 206, 484, 555 ], "group_id": 0, "lines": [ { "bbox": [ 132, 206, 484, 555 ], "spans": [ { "bbox": [ 132, 206, 484, 555 ], "score": 0.972, "type": "image", "image_path": "8f5416bea6a57fe53694b7b81ca32fab188b10b5307dd5743801f95cfe3dc61b.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 206, 484, 322.3333333333333 ], "spans": [], "index": 0 }, { "bbox": [ 132, 322.3333333333333, 484, 438.66666666666663 ], "spans": [], "index": 1 }, { "bbox": [ 132, 438.66666666666663, 484, 555.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 105, 567, 505, 579 ], "spans": [ { "bbox": [ 105, 567, 505, 579 ], "score": 1.0, "content": "Figure 23: Reference-guided image synthesis results DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 578, 505, 590 ], "spans": [ { "bbox": [ 105, 578, 505, 590 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 589, 505, 601 ], "spans": [ { "bbox": [ 105, 589, 505, 601 ], "score": 1.0, "content": "from Normal to Scratches by taking the Source as background and applying the foreground", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 600, 221, 612 ], "spans": [ { "bbox": [ 106, 600, 221, 612 ], "score": 1.0, "content": "defect from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ], "page_idx": 35, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 763 ], "spans": [ { "bbox": [ 298, 750, 313, 763 ], "score": 1.0, "content": "36", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 206, 484, 555 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 206, 484, 555 ], "group_id": 0, "lines": [ { "bbox": [ 132, 206, 484, 555 ], "spans": [ { "bbox": [ 132, 206, 484, 555 ], "score": 0.972, "type": "image", "image_path": "8f5416bea6a57fe53694b7b81ca32fab188b10b5307dd5743801f95cfe3dc61b.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 206, 484, 322.3333333333333 ], "spans": [], "index": 0 }, { "bbox": [ 132, 322.3333333333333, 484, 438.66666666666663 ], "spans": [], "index": 1 }, { "bbox": [ 132, 438.66666666666663, 484, 555.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 105, 567, 505, 579 ], "spans": [ { "bbox": [ 105, 567, 505, 579 ], "score": 1.0, "content": "Figure 23: Reference-guided image synthesis results DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 578, 505, 590 ], "spans": [ { "bbox": [ 105, 578, 505, 590 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 589, 505, 601 ], "spans": [ { "bbox": [ 105, 589, 505, 601 ], "score": 1.0, "content": "from Normal to Scratches by taking the Source as background and applying the foreground", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 600, 221, 612 ], "spans": [ { "bbox": [ 106, 600, 221, 612 ], "score": 1.0, "content": "defect from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 132, 208, 484, 555 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 208, 484, 555 ], "group_id": 0, "lines": [ { "bbox": [ 132, 208, 484, 555 ], "spans": [ { "bbox": [ 132, 208, 484, 555 ], "score": 0.969, "type": "image", "image_path": "038273ee81067405ed8edb30707628ce21e36f0f73845fa8004a02f19865a69d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 208, 484, 323.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 132, 323.6666666666667, 484, 439.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 132, 439.33333333333337, 484, 555.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 106, 566, 506, 579 ], "spans": [ { "bbox": [ 106, 566, 506, 579 ], "score": 1.0, "content": "Figure 24: Reference-guided image synthesis results of DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 578, 505, 590 ], "spans": [ { "bbox": [ 105, 578, 505, 590 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 505, 602 ], "score": 1.0, "content": "from Normal to Spots by taking the Source as background and applying the foreground defect", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 600, 194, 612 ], "spans": [ { "bbox": [ 106, 600, 194, 612 ], "score": 1.0, "content": "from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ], "page_idx": 36, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 308, 37 ], "lines": [ { "bbox": [ 106, 26, 308, 38 ], "spans": [ { "bbox": [ 106, 26, 308, 38 ], "score": 1.0, "content": "Under review as a conference paper at ICLR 2022", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 132, 208, 484, 555 ], "blocks": [ { "type": "image_body", "bbox": [ 132, 208, 484, 555 ], "group_id": 0, "lines": [ { "bbox": [ 132, 208, 484, 555 ], "spans": [ { "bbox": [ 132, 208, 484, 555 ], "score": 0.969, "type": "image", "image_path": "038273ee81067405ed8edb30707628ce21e36f0f73845fa8004a02f19865a69d.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 132, 208, 484, 323.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 132, 323.6666666666667, 484, 439.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 132, 439.33333333333337, 484, 555.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 567, 506, 612 ], "group_id": 0, "lines": [ { "bbox": [ 106, 566, 506, 579 ], "spans": [ { "bbox": [ 106, 566, 506, 579 ], "score": 1.0, "content": "Figure 24: Reference-guided image synthesis results of DT-GAN on the SDI dataset and the MVTec", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 578, 505, 590 ], "spans": [ { "bbox": [ 105, 578, 505, 590 ], "score": 1.0, "content": "AD dataset. The model is trained on a joint set of aforementioned datasets and performs translation", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 505, 602 ], "score": 1.0, "content": "from Normal to Spots by taking the Source as background and applying the foreground defect", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 600, 194, 612 ], "spans": [ { "bbox": [ 106, 600, 194, 612 ], "score": 1.0, "content": "from Reference to it.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 } ] } ], "_backend": "pipeline", "_version_name": "2.2.2" }