| Task | Tools | Model Description |
| Super-resolution | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Denoising | SCUnet [94] | Hybrid UNet-based model combining convolution and transformer blocks, designed for robust denoising under diverse real-world noise conditions. |
| Compression artifact removal | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Deblurring | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently.. |
| Deraining | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| UDR-S2Former [8] | An uncertainty-aware transformer model for rain streak removal. |
| Img2img-turbo-rain [52] | Efficient model based on SD-turbo, designed for fast and effective rain removal in real-world images. |
| Raindrop removal | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| Dehazing | RIDCP [81] | Efficient dehazing model utilizing high-quality codebook priors to handle complex real-world haze. |
| KANet [21] | Efficient dehazing network using a localization-and-removal pipeline to handle complex real-world hazy. |
| Desnowing | Img2img-turbo-snow [52] | Efficient model for removing snow artifacts while preserving natural scene details. |
| Snowformer [11] | Transformer-based model for removing snowflakes while preserving natural scene details. |
| Low-light enhancement | Img2img-turbo-night [52] | Fast and efficient model based on SD-turbo, designed for low-light enhancement in real-world scenarios. |
| HVI-CIDNet [83] | Lightweight transformer for low-light and exposure correction, enhancing both image quality and downstream vision tasks efficiently. |
| LightenDiff [27] | Diffusion-based framework for low-light enhancement, leveraging Retinex theory and latent-space decomposition for high-quality unsupervised restoration. |
",
+ "bbox": [
+ 96,
+ 114,
+ 901,
+ 760
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "come from diverse sources, including the aforementioned autonomous driving datasets. Additionally, to enhance the generalizability of JarvisIR in natural contexts, we incorporated natural adverse weather scenes from internet and public datasets [5, 34, 39, 42, 47, 57, 87, 99].",
+ "bbox": [
+ 89,
+ 784,
+ 482,
+ 859
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "9.2. Details of degradation library",
+ "text_level": 1,
+ "bbox": [
+ 513,
+ 782,
+ 777,
+ 800
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "As described in Sec 3.1 of the manuscript, we simulate realistic adverse weather scenarios—rainy, nighttime, snowy, and foggy conditions—by customizing a degradation library developed with physical models and image transformation techniques to synthesize degraded images. In this section, we detail our degradation implementations, cov",
+ "bbox": [
+ 511,
+ 810,
+ 906,
+ 902
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "page_number",
+ "text": "15",
+ "bbox": [
+ 490,
+ 924,
+ 508,
+ 936
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "image",
+ "img_path": "images/71a63f46812d2f4fe58ba8bc6a43938249dabe1c45d6bd32c8ceaa33a43536fc.jpg",
+ "image_caption": [
+ "Figure 8. Adverse weather scene simulator. To simulate realistic adverse weather scenarios, including rainy, nighttime, snowy, and foggy, we customized the degradation library developed using physical models and image transformation techniques to synthesize degraded images."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 96,
+ 98,
+ 890,
+ 568
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "ering the principles, formulas, and severity setups for the Night Scene Simulator, Fog Scene Simulator, Rain Scene Simulator, and Snow Scene Simulator. Examples for each implementation are provided in Figure 11.",
+ "bbox": [
+ 89,
+ 637,
+ 486,
+ 699
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Night Scene Simulator. Inspired by the work of [19], we employ a low-light degradation transform to synthesize realistic low-light images, denoted as $T_{\\text{night}}$ , as illustrated in Figure 8. Specifically, we first convert the daytime image $I_{\\text{day}}$ into RAW data using the sRGB $\\rightarrow$ RAW process [3]. Next, we linearly attenuate the RAW image and introduce Shot and Read (S&R) noise, which is commonly observed in camera imaging systems [3]. Finally, we apply the Image Signal Processing (ISP) pipeline to convert the low-light sensor data back into sRGB format. Additionally, we incorporate flare degradation using flare templates from the Flare7K++ [20] dataset. The complete low-light degrada",
+ "bbox": [
+ 89,
+ 719,
+ 486,
+ 902
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "tion transform $T_{\\text{night}}$ is given by:",
+ "bbox": [
+ 511,
+ 638,
+ 740,
+ 655
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nT _ {\\text {n i g h t}} \\left(\\mathrm {I} _ {\\text {d a y}}\\right) = T _ {I S P} \\left(T _ {s R G B \\rightarrow R A W} \\left(\\mathrm {I} _ {\\text {d a y}}\\right) + \\mathrm {I} _ {\\text {n o i s e}}\\right) + \\mathrm {I} _ {\\text {f l a r e}}, \\tag {8}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 511,
+ 670,
+ 911,
+ 700
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "which generates a degraded image $I_{day}$ that closely resembles a dark nighttime scene. Furthermore, we use an online dynamic degradation process. It applies randomized parameter combinations, as defined in Equation 8, to simulate diverse nighttime driving conditions.",
+ "bbox": [
+ 511,
+ 700,
+ 906,
+ 777
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "Fog Scene Simulator. Inspired by RIDCP [81], we design a foggy image degradation transform, denoted as $T_{\\mathrm{fog}}$ , to synthesize realistic hazy images, as shown in Figure 8. Specifically, we simulate fog by introducing transmission maps $t(x)$ using depth estimation algorithms (e.g., Depth anything V2 [84]), combined with exponential attenuation $e^{\\beta d(x)}$ , where $\\beta$ controls haze density within the range [0.3, 1.5]. Additionally, poor lighting conditions are",
+ "bbox": [
+ 511,
+ 779,
+ 908,
+ 902
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "16",
+ "bbox": [
+ 490,
+ 924,
+ 509,
+ 936
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "text",
+ "text": "modeled by applying a brightness adjustment factor $\\gamma \\in [1.5,3.0]$ , Gaussian noise $\\mathcal{N}$ , and atmospheric light variation $A + \\Delta A$ , where $\\Delta A$ is sampled from $[-0.025,0.025]$ . To further enhance realism, JPEG compression artifacts are introduced by applying JPEG $(\\cdot)$ to the degraded image. The complete foggy image synthesis process is defined as:",
+ "bbox": [
+ 89,
+ 90,
+ 480,
+ 183
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nT _ {\\text {f o g}} \\left(I _ {\\text {d a y}}\\right) = \\operatorname {J P E G} \\left(\\mathcal {P} \\left(I _ {\\text {d a y}} ^ {\\gamma} + \\mathcal {N}, e ^ {\\beta d (x)}, A + \\Delta A\\right)\\right), \\tag {9}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 94,
+ 186,
+ 482,
+ 224
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathcal{P}$ represents the hazy image formation process, $I_{\\mathrm{day}}$ is the clean image, and $d(x)$ is the estimated depth map. The variable $x$ refers to the spatial coordinates of the image. This dynamic degradation process is designed to operate online with randomized parameters, simulating diverse real-world foggy conditions.",
+ "bbox": [
+ 89,
+ 224,
+ 480,
+ 314
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Rain Scene Simulator. Inspired by PGDGN [54], we introduce a rain degradation transform, denoted as $T_{\\mathrm{rain}}$ , to generate realistic rainy images (Figure 8). This transform synthesizes rainy images by combining a disentangled clean image with a physics-based rain rendering model. The degradation process is formulated as:",
+ "bbox": [
+ 89,
+ 315,
+ 480,
+ 405
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nT _ {\\text {r a i n}} \\left(I _ {\\text {d a y}}\\right) = W _ {\\text {M o d}} \\left(G \\left(I _ {\\text {d a y}}\\right), \\hat {w}, z\\right), \\tag {10}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 158,
+ 412,
+ 480,
+ 429
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "where $I_{\\mathrm{day}}$ is the clean image, $G(I_{\\mathrm{day}})$ represents the disentangled base image, and $W_{\\mathrm{Mod}}$ is the rain rendering model. $W_{\\mathrm{Mod}}$ incorporates parameters $\\hat{w} = \\{\\hat{w}_d, \\hat{w}_{nd}\\}$ , with $\\hat{w}_d$ controlling differentiable aspects such as raindrop size and streak density, and $\\hat{w}_{nd}$ addressing nondifferentiable properties. The term $z$ introduces stochastic noise for variability in rain effects. This process applies $W_{\\mathrm{Mod}}$ to add realistic raindrop occlusions, rain streaks, and scene wetness to the disentangled image. $G(I_{\\mathrm{day}})$ , generating a visually plausible rainy image $T_{\\mathrm{rain}}(I_{\\mathrm{day}})$ with controlled and diverse effects.",
+ "bbox": [
+ 89,
+ 434,
+ 482,
+ 598
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "Snow Scene Simulator. Building on the img2img-turbo model [52], we introduce a snow transformation, denoted as $T_{\\mathrm{snow}}$ , to generate realistic snowy images from daytime inputs. This process uses the SD-Turbo model with textual conditioning. It synthesizes snowy scenes by combining the input image with a latent diffusion-based generator and a textual prompt. The snow transformation is formulated as:",
+ "bbox": [
+ 89,
+ 599,
+ 482,
+ 705
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nT _ {\\text {s n o w}} \\left(I _ {\\text {d a y}}, C _ {\\text {s n o w}}\\right) = G _ {\\text {s n o w}} \\left(I _ {\\text {d a y}}, C _ {\\text {s n o w}}\\right), \\tag {11}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 127,
+ 712,
+ 480,
+ 729
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "where $I_{\\mathrm{day}}$ is the daytime input image, $C_{\\mathrm{snow}}$ is the textual condition (e.g., \"driving in the heavy snow\"), and $G_{\\mathrm{snow}}$ represents the generator. By employing LoRA adapters and skip connections, the generator enables precise control over scene characteristics while maintaining the structural integrity of the input image. This process applies $G_{\\mathrm{snow}}$ to infuse the daytime image $I_{\\mathrm{day}}$ with snowy features, guided by the contextual information in $C_{\\mathrm{snow}}$ . The resulting synthetic image aligns closely with the visual expectations of a snowy environment while maintaining consistency with the original scene's structure.",
+ "bbox": [
+ 89,
+ 734,
+ 482,
+ 900
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "10. More ablation",
+ "text_level": 1,
+ "bbox": [
+ 514,
+ 89,
+ 668,
+ 104
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "To thoroughly investigate the proposed JarvisIR, we conducted an extensive array of ablation studies on the CleanBench-Real dataset. Four non-reference metrics are used for assessment: MUSIQ [35], MANIQA [86], CLIP-IQA+ [69], LIQE [95]. The specific elements of these studies are further expounded in the sections that follow.",
+ "bbox": [
+ 511,
+ 114,
+ 903,
+ 205
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "10.1. MRRHF vs. vanilla RRHF",
+ "text_level": 1,
+ "bbox": [
+ 513,
+ 217,
+ 764,
+ 232
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "We evaluate the effectiveness of our proposed MRRHF by comparing it with vanilla RRHF [93]. The reward and diversity metrics over training iterations are illustrated in Table 5. Fine-tuning JarvisIR with MRRHF significantly improves the average values of both reward and diversity by 0.19 and 3.43, respectively, compared to using RRHF. The degradation in diversity and reward when using vanilla RRHF results from its offline sample generation strategy. As discussed in Sec. 5 in the manuscript, this strategy confines its generated samples to the finite sample space created by the SFT model using diverse beam search [68]. In contrast, our MRRHF employs a hybrid sample generation strategy and entropy regularization, providing sufficient sample exploration space to achieve globally optimal results.",
+ "bbox": [
+ 511,
+ 239,
+ 906,
+ 467
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "10.2. Sample generation strategy and entropy regularization",
+ "text_level": 1,
+ "bbox": [
+ 514,
+ 479,
+ 903,
+ 508
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "In our manuscript, we examine the effects of the sample generation strategy and entropy regularization on the MR-RHF tuning process, focusing on reward scores and response diversity. This section provides further evidence of the effectiveness of our hybrid sample generation strategy and entropy regularization. Specifically, as shown in Table 5, we assess their impact on performance using the CleanBench-Real validation set. The results demonstrate that our hybrid sampling approach and entropy regularization not only enhance training stability and facilitate high-quality exploration of the optimization space but also significantly improve testing performance.",
+ "bbox": [
+ 511,
+ 516,
+ 905,
+ 699
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "10.3. Effectiveness of differentiated contrast weights",
+ "text_level": 1,
+ "bbox": [
+ 513,
+ 709,
+ 903,
+ 741
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "text",
+ "text": "In Equation 4 of our manuscript, we refine the original ranking loss [93] by introducing differentiated contrast weights, expressed as $L_{\\mathrm{rank}} = \\sum_{s_i < s_j} (s_j - s_i) \\max(0, p_i - p_j)$ . The term $(s_j - s_i)$ represents the differentiated contrast weights. We compare this with the original ranking loss $\\hat{L}_{\\mathrm{rank}} = \\sum_{s_i < s_j} \\max(0, p_i - p_j)$ . Table 5 presents the reward and diversity metrics over training iterations. When the ranking loss is applied without differentiated contrast weights $\\hat{L}_{\\mathrm{rank}}$ the average values of both reward and diversity decrease by 0.14 and 3.93, respectively, compared",
+ "bbox": [
+ 511,
+ 747,
+ 906,
+ 901
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "page_number",
+ "text": "17",
+ "bbox": [
+ 490,
+ 924,
+ 508,
+ 936
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "table",
+ "img_path": "images/048f68f8d200027e9e905194273c996241437124d691fadea0d3323580d729c0.jpg",
+ "table_caption": [
+ "Table 5. Ablation studies on tuning paradigm, differentiated contrast weights, different sample generation strategies, and entropy regularization. The \"Reward\" represents the average reward scores obtained during alignment tuning, spanning from -1 to 1. A negative score indicates a penalty, while a positive score represents a reward. The \"Diversity\" reflects the average number of unique responses produced during the training process. Additionally, we evaluate performance on the CleanBench-Real validation set using four non-reference metrics: MUSIQ, MANIQA, CLIP-IQA+, and LIQE. The reported values represent the average performance across all tested scenes."
+ ],
+ "table_footnote": [],
+ "table_body": "| # | Instruction |
| 1 | Please evaluate the image's quality comprehensively and provide your insights. Additionally, outline a step-by-step restoration strategy, specifying the sequence of tasks and model choices for each step. The available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement. The available restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 2 | Analyze the quality of the image comprehensively and provide your insights. Furthermore, propose a restoration strategy by detailing each task and model choice sequentially. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement. The restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 3 | Assess the overall quality of the image and provide a detailed evaluation. Then, design a step-by-step restoration process, specifying tasks and model choices. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 4 | Perform a comprehensive evaluation of the image quality and explain your observations. Additionally, develop a step-by-step restoration plan, identifying tasks and model choices. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 5 | Conduct a thorough analysis of the image's quality and provide your insights. Subsequently, create a restoration strategy step by step, specifying the tasks and model choices. The tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 6 | Evaluate the quality of the image comprehensively and outline your findings. Moreover, formulate a sequential restoration plan, detailing tasks and model selections. Available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 7 | Provide a detailed assessment of the image's quality and share your observations. Then, create a restoration strategy in a step-by-step manner, specifying tasks and models. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 8 | Examine the image comprehensively and describe its quality in detail. Additionally, outline a step-wise restoration process, specifying the tasks and models for each stage. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 9 | Review the image quality comprehensively and present your findings. Furthermore, propose a restoration plan step by step, detailing the tasks and models used. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 10 | Perform a detailed evaluation of the image and provide your insights. Then, construct a step-by-step restoration strategy, specifying the sequence of tasks and the corresponding models. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
",
+ "bbox": [
+ 127,
+ 155,
+ 867,
+ 858
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "page_number",
+ "text": "24",
+ "bbox": [
+ 488,
+ 924,
+ 508,
+ 936
+ ],
+ "page_idx": 23
+ },
+ {
+ "type": "table",
+ "img_path": "images/46e0fd7cfb79df6a4f73542600a61049ba8364fa75e365b59aaf3d2e9e3a372c.jpg",
+ "table_caption": [
+ "Table 9. Responses generated by GPT-4V using the self-instruct strategy [73]"
+ ],
+ "table_footnote": [],
+ "table_body": "| # | Response |
| 1 | The image depicts an outdoor scene viewed through a wet windshield, with water droplets significantly obscuring the view. The evaluated image exhibits severe distortion caused by the water droplets, resulting in a loss of sharpness and clarity. Details of the background, such as the car ahead and surrounding structures, are blurred and difficult to discern. The image also suffers from uneven focus and lacks proper visibility. |
| 2 | This image portrays an outdoor scene as seen through a wet windshield, where water droplets heavily obstruct the view. The distortion caused by these droplets leads to a noticeable reduction in sharpness and clarity. Background elements, like the car in front and nearby structures, appear heavily blurred and challenging to distinguish. The image further suffers from uneven focus, reducing overall visibility. |
| 3 | An outdoor setting is captured in this image, viewed through a wet windshield. The water droplets significantly obscure the scene, causing severe distortion. Sharpness and clarity are greatly reduced, and background features, such as the car ahead and surrounding buildings, are blurred and hard to recognize. Uneven focus further detracts from the image's overall visibility. |
| 4 | This image showcases an outdoor scene as observed through a wet windshield, with water droplets causing significant obstruction. The distortion leads to substantial clarity and sharpness loss. Details in the background, including the car ahead and nearby structures, are blurred and hard to identify. Additionally, the image has uneven focus, which impairs its overall visibility. |
| 5 | The image illustrates an outdoor view observed through a wet windshield, where water droplets obscure much of the scene. The distortion caused by the droplets results in a notable reduction in sharpness and clarity. Background details, such as the car ahead and surrounding structures, appear blurry and indistinct. Uneven focus further degrades the visibility of the image. |
| 6 | In this image, an outdoor setting is depicted as seen through a wet windshield, with water droplets heavily obscuring the view. The distortion from the droplets results in a considerable loss of sharpness and clarity. Background elements, like the car ahead and nearby buildings, are blurred and hard to discern. The uneven focus of the image adds to its overall lack of visibility. |
| 7 | This image shows an outdoor scene viewed through a wet windshield, with water droplets creating significant visual obstruction. The distortion caused by the droplets reduces the sharpness and clarity of the image. Background features, including the car ahead and the surrounding structures, are blurred and difficult to distinguish. The image also suffers from uneven focus, limiting its visibility. |
| 8 | An outdoor view is displayed in this image, captured through a wet windshield with water droplets blocking much of the scene. The distortion caused by the droplets results in reduced clarity and sharpness. Background details, such as the car in front and nearby structures, are blurred and indistinct. Uneven focus further compromises the visibility of the image. |
| 9 | The image depicts an outdoor scene seen through a wet windshield, where water droplets obscure much of the view. This distortion leads to a significant loss of sharpness and clarity. Background elements like the car ahead and nearby structures are blurred and challenging to distinguish. Uneven focus further contributes to the poor visibility of the image. |
| 10 | This image represents an outdoor scene viewed through a wet windshield, with water droplets obscuring the visual field. The distortion caused by the droplets significantly affects sharpness and clarity, making background features like the car ahead and surrounding structures appear blurred and indistinct. The uneven focus further reduces the overall visibility of the image. |
",
+ "bbox": [
+ 127,
+ 148,
+ 867,
+ 864
+ ],
+ "page_idx": 24
+ },
+ {
+ "type": "page_number",
+ "text": "25",
+ "bbox": [
+ 488,
+ 924,
+ 506,
+ 936
+ ],
+ "page_idx": 24
+ }
+]
\ No newline at end of file
diff --git a/data/2025/2504_04xxx/2504.04158/f2103d47-786d-466a-b93a-76660f3198b5_model.json b/data/2025/2504_04xxx/2504.04158/f2103d47-786d-466a-b93a-76660f3198b5_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..936f043f23cc2c9c5a4bd73e680a884a8f341ad6
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/f2103d47-786d-466a-b93a-76660f3198b5_model.json
@@ -0,0 +1,4903 @@
+[
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.21,
+ 0.124,
+ 0.258,
+ 0.16
+ ],
+ "angle": 0,
+ "content": "#"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.262,
+ 0.139,
+ 0.79,
+ 0.184
+ ],
+ "angle": 0,
+ "content": "JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.176,
+ 0.211,
+ 0.825,
+ 0.229
+ ],
+ "angle": 0,
+ "content": "Yunlong \\(\\mathrm{Lin}^{1*}\\) Zixu Lin\\(^{1*}\\) Haoyu Chen\\(^{2*}\\) Panwang Pan\\(^{3*}\\) Chenxin Li\\(^{6}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.259,
+ 0.23,
+ 0.756,
+ 0.248
+ ],
+ "angle": 0,
+ "content": "Sixiang Chen² Yeying Jin⁴ Wenbo Li⁵† Xinghao Ding¹"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.194,
+ 0.248,
+ 0.805,
+ 0.266
+ ],
+ "angle": 0,
+ "content": "| Task | Tools | Model Description |
| Super-resolution | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Denoising | SCUnet [94] | Hybrid UNet-based model combining convolution and transformer blocks, designed for robust denoising under diverse real-world noise conditions. |
| Compression artifact removal | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Deblurring | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently.. |
| Deraining | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| UDR-S2Former [8] | An uncertainty-aware transformer model for rain streak removal. |
| Img2img-turbo-rain [52] | Efficient model based on SD-turbo, designed for fast and effective rain removal in real-world images. |
| Raindrop removal | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| Dehazing | RIDCP [81] | Efficient dehazing model utilizing high-quality codebook priors to handle complex real-world haze. |
| KANet [21] | Efficient dehazing network using a localization-and-removal pipeline to handle complex real-world hazy. |
| Desnowing | Img2img-turbo-snow [52] | Efficient model for removing snow artifacts while preserving natural scene details. |
| Snowformer [11] | Transformer-based model for removing snowflakes while preserving natural scene details. |
| Low-light enhancement | Img2img-turbo-night [52] | Fast and efficient model based on SD-turbo, designed for low-light enhancement in real-world scenarios. |
| HVI-CIDNet [83] | Lightweight transformer for low-light and exposure correction, enhancing both image quality and downstream vision tasks efficiently. |
| LightenDiff [27] | Diffusion-based framework for low-light enhancement, leveraging Retinex theory and latent-space decomposition for high-quality unsupervised restoration. |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.785,
+ 0.483,
+ 0.86
+ ],
+ "angle": 0,
+ "content": "come from diverse sources, including the aforementioned autonomous driving datasets. Additionally, to enhance the generalizability of JarvisIR in natural contexts, we incorporated natural adverse weather scenes from internet and public datasets [5, 34, 39, 42, 47, 57, 87, 99]."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.514,
+ 0.784,
+ 0.779,
+ 0.801
+ ],
+ "angle": 0,
+ "content": "9.2. Details of degradation library"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.811,
+ 0.907,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "As described in Sec 3.1 of the manuscript, we simulate realistic adverse weather scenarios—rainy, nighttime, snowy, and foggy conditions—by customizing a degradation library developed with physical models and image transformation techniques to synthesize degraded images. In this section, we detail our degradation implementations, cov"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.925,
+ 0.509,
+ 0.937
+ ],
+ "angle": 0,
+ "content": "15"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.097,
+ 0.099,
+ 0.891,
+ 0.569
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.09,
+ 0.58,
+ 0.908,
+ 0.625
+ ],
+ "angle": 0,
+ "content": "Figure 8. Adverse weather scene simulator. To simulate realistic adverse weather scenarios, including rainy, nighttime, snowy, and foggy, we customized the degradation library developed using physical models and image transformation techniques to synthesize degraded images."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.638,
+ 0.487,
+ 0.7
+ ],
+ "angle": 0,
+ "content": "ering the principles, formulas, and severity setups for the Night Scene Simulator, Fog Scene Simulator, Rain Scene Simulator, and Snow Scene Simulator. Examples for each implementation are provided in Figure 11."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.72,
+ 0.487,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "Night Scene Simulator. Inspired by the work of [19], we employ a low-light degradation transform to synthesize realistic low-light images, denoted as \\( T_{\\text{night}} \\), as illustrated in Figure 8. Specifically, we first convert the daytime image \\( I_{\\text{day}} \\) into RAW data using the sRGB \\(\\rightarrow\\) RAW process [3]. Next, we linearly attenuate the RAW image and introduce Shot and Read (S&R) noise, which is commonly observed in camera imaging systems [3]. Finally, we apply the Image Signal Processing (ISP) pipeline to convert the low-light sensor data back into sRGB format. Additionally, we incorporate flare degradation using flare templates from the Flare7K++ [20] dataset. The complete low-light degrada"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.513,
+ 0.639,
+ 0.741,
+ 0.656
+ ],
+ "angle": 0,
+ "content": "tion transform \\(T_{\\text{night}}\\) is given by:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.513,
+ 0.671,
+ 0.913,
+ 0.702
+ ],
+ "angle": 0,
+ "content": "\\[\nT _ {\\text {n i g h t}} \\left(\\mathrm {I} _ {\\text {d a y}}\\right) = T _ {I S P} \\left(T _ {s R G B \\rightarrow R A W} \\left(\\mathrm {I} _ {\\text {d a y}}\\right) + \\mathrm {I} _ {\\text {n o i s e}}\\right) + \\mathrm {I} _ {\\text {f l a r e}}, \\tag {8}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.702,
+ 0.907,
+ 0.778
+ ],
+ "angle": 0,
+ "content": "which generates a degraded image \\( I_{day} \\) that closely resembles a dark nighttime scene. Furthermore, we use an online dynamic degradation process. It applies randomized parameter combinations, as defined in Equation 8, to simulate diverse nighttime driving conditions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.78,
+ 0.909,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "Fog Scene Simulator. Inspired by RIDCP [81], we design a foggy image degradation transform, denoted as \\( T_{\\mathrm{fog}} \\), to synthesize realistic hazy images, as shown in Figure 8. Specifically, we simulate fog by introducing transmission maps \\( t(x) \\) using depth estimation algorithms (e.g., Depth anything V2 [84]), combined with exponential attenuation \\( e^{\\beta d(x)} \\), where \\( \\beta \\) controls haze density within the range [0.3, 1.5]. Additionally, poor lighting conditions are"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.925,
+ 0.51,
+ 0.937
+ ],
+ "angle": 0,
+ "content": "16"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.092,
+ 0.482,
+ 0.184
+ ],
+ "angle": 0,
+ "content": "modeled by applying a brightness adjustment factor \\(\\gamma \\in [1.5,3.0]\\), Gaussian noise \\(\\mathcal{N}\\), and atmospheric light variation \\(A + \\Delta A\\), where \\(\\Delta A\\) is sampled from \\([-0.025,0.025]\\). To further enhance realism, JPEG compression artifacts are introduced by applying JPEG \\((\\cdot)\\) to the degraded image. The complete foggy image synthesis process is defined as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.095,
+ 0.187,
+ 0.483,
+ 0.225
+ ],
+ "angle": 0,
+ "content": "\\[\nT _ {\\text {f o g}} \\left(I _ {\\text {d a y}}\\right) = \\operatorname {J P E G} \\left(\\mathcal {P} \\left(I _ {\\text {d a y}} ^ {\\gamma} + \\mathcal {N}, e ^ {\\beta d (x)}, A + \\Delta A\\right)\\right), \\tag {9}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.226,
+ 0.482,
+ 0.315
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathcal{P}\\) represents the hazy image formation process, \\(I_{\\mathrm{day}}\\) is the clean image, and \\(d(x)\\) is the estimated depth map. The variable \\(x\\) refers to the spatial coordinates of the image. This dynamic degradation process is designed to operate online with randomized parameters, simulating diverse real-world foggy conditions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.316,
+ 0.482,
+ 0.406
+ ],
+ "angle": 0,
+ "content": "Rain Scene Simulator. Inspired by PGDGN [54], we introduce a rain degradation transform, denoted as \\( T_{\\mathrm{rain}} \\), to generate realistic rainy images (Figure 8). This transform synthesizes rainy images by combining a disentangled clean image with a physics-based rain rendering model. The degradation process is formulated as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.16,
+ 0.413,
+ 0.482,
+ 0.43
+ ],
+ "angle": 0,
+ "content": "\\[\nT _ {\\text {r a i n}} \\left(I _ {\\text {d a y}}\\right) = W _ {\\text {M o d}} \\left(G \\left(I _ {\\text {d a y}}\\right), \\hat {w}, z\\right), \\tag {10}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.435,
+ 0.483,
+ 0.599
+ ],
+ "angle": 0,
+ "content": "where \\( I_{\\mathrm{day}} \\) is the clean image, \\( G(I_{\\mathrm{day}}) \\) represents the disentangled base image, and \\( W_{\\mathrm{Mod}} \\) is the rain rendering model. \\( W_{\\mathrm{Mod}} \\) incorporates parameters \\( \\hat{w} = \\{\\hat{w}_d, \\hat{w}_{nd}\\} \\), with \\( \\hat{w}_d \\) controlling differentiable aspects such as raindrop size and streak density, and \\( \\hat{w}_{nd} \\) addressing nondifferentiable properties. The term \\( z \\) introduces stochastic noise for variability in rain effects. This process applies \\( W_{\\mathrm{Mod}} \\) to add realistic raindrop occlusions, rain streaks, and scene wetness to the disentangled image. \\( G(I_{\\mathrm{day}}) \\), generating a visually plausible rainy image \\( T_{\\mathrm{rain}}(I_{\\mathrm{day}}) \\) with controlled and diverse effects."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.601,
+ 0.483,
+ 0.707
+ ],
+ "angle": 0,
+ "content": "Snow Scene Simulator. Building on the img2img-turbo model [52], we introduce a snow transformation, denoted as \\( T_{\\mathrm{snow}} \\), to generate realistic snowy images from daytime inputs. This process uses the SD-Turbo model with textual conditioning. It synthesizes snowy scenes by combining the input image with a latent diffusion-based generator and a textual prompt. The snow transformation is formulated as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.128,
+ 0.713,
+ 0.482,
+ 0.73
+ ],
+ "angle": 0,
+ "content": "\\[\nT _ {\\text {s n o w}} \\left(I _ {\\text {d a y}}, C _ {\\text {s n o w}}\\right) = G _ {\\text {s n o w}} \\left(I _ {\\text {d a y}}, C _ {\\text {s n o w}}\\right), \\tag {11}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.735,
+ 0.483,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "where \\( I_{\\mathrm{day}} \\) is the daytime input image, \\( C_{\\mathrm{snow}} \\) is the textual condition (e.g., \"driving in the heavy snow\"), and \\( G_{\\mathrm{snow}} \\) represents the generator. By employing LoRA adapters and skip connections, the generator enables precise control over scene characteristics while maintaining the structural integrity of the input image. This process applies \\( G_{\\mathrm{snow}} \\) to infuse the daytime image \\( I_{\\mathrm{day}} \\) with snowy features, guided by the contextual information in \\( C_{\\mathrm{snow}} \\). The resulting synthetic image aligns closely with the visual expectations of a snowy environment while maintaining consistency with the original scene's structure."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.09,
+ 0.669,
+ 0.106
+ ],
+ "angle": 0,
+ "content": "10. More ablation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.116,
+ 0.905,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "To thoroughly investigate the proposed JarvisIR, we conducted an extensive array of ablation studies on the CleanBench-Real dataset. Four non-reference metrics are used for assessment: MUSIQ [35], MANIQA [86], CLIP-IQA+ [69], LIQE [95]. The specific elements of these studies are further expounded in the sections that follow."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.514,
+ 0.218,
+ 0.766,
+ 0.233
+ ],
+ "angle": 0,
+ "content": "10.1. MRRHF vs. vanilla RRHF"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.241,
+ 0.907,
+ 0.468
+ ],
+ "angle": 0,
+ "content": "We evaluate the effectiveness of our proposed MRRHF by comparing it with vanilla RRHF [93]. The reward and diversity metrics over training iterations are illustrated in Table 5. Fine-tuning JarvisIR with MRRHF significantly improves the average values of both reward and diversity by 0.19 and 3.43, respectively, compared to using RRHF. The degradation in diversity and reward when using vanilla RRHF results from its offline sample generation strategy. As discussed in Sec. 5 in the manuscript, this strategy confines its generated samples to the finite sample space created by the SFT model using diverse beam search [68]. In contrast, our MRRHF employs a hybrid sample generation strategy and entropy regularization, providing sufficient sample exploration space to achieve globally optimal results."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.515,
+ 0.48,
+ 0.905,
+ 0.51
+ ],
+ "angle": 0,
+ "content": "10.2. Sample generation strategy and entropy regularization"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.517,
+ 0.906,
+ 0.7
+ ],
+ "angle": 0,
+ "content": "In our manuscript, we examine the effects of the sample generation strategy and entropy regularization on the MR-RHF tuning process, focusing on reward scores and response diversity. This section provides further evidence of the effectiveness of our hybrid sample generation strategy and entropy regularization. Specifically, as shown in Table 5, we assess their impact on performance using the CleanBench-Real validation set. The results demonstrate that our hybrid sampling approach and entropy regularization not only enhance training stability and facilitate high-quality exploration of the optimization space but also significantly improve testing performance."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.514,
+ 0.71,
+ 0.905,
+ 0.742
+ ],
+ "angle": 0,
+ "content": "10.3. Effectiveness of differentiated contrast weights"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.512,
+ 0.748,
+ 0.907,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "In Equation 4 of our manuscript, we refine the original ranking loss [93] by introducing differentiated contrast weights, expressed as \\( L_{\\mathrm{rank}} = \\sum_{s_i < s_j} (s_j - s_i) \\max(0, p_i - p_j) \\). The term \\( (s_j - s_i) \\) represents the differentiated contrast weights. We compare this with the original ranking loss \\( \\hat{L}_{\\mathrm{rank}} = \\sum_{s_i < s_j} \\max(0, p_i - p_j) \\). Table 5 presents the reward and diversity metrics over training iterations. When the ranking loss is applied without differentiated contrast weights \\( \\hat{L}_{\\mathrm{rank}} \\) the average values of both reward and diversity decrease by 0.14 and 3.93, respectively, compared"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.491,
+ 0.925,
+ 0.509,
+ 0.937
+ ],
+ "angle": 0,
+ "content": "17"
+ }
+ ],
+ [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.09,
+ 0.085,
+ 0.907,
+ 0.157
+ ],
+ "angle": 0,
+ "content": "Table 5. Ablation studies on tuning paradigm, differentiated contrast weights, different sample generation strategies, and entropy regularization. The \"Reward\" represents the average reward scores obtained during alignment tuning, spanning from -1 to 1. A negative score indicates a penalty, while a positive score represents a reward. The \"Diversity\" reflects the average number of unique responses produced during the training process. Additionally, we evaluate performance on the CleanBench-Real validation set using four non-reference metrics: MUSIQ, MANIQA, CLIP-IQA+, and LIQE. The reported values represent the average performance across all tested scenes."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.113,
+ 0.158,
+ 0.895,
+ 0.34
+ ],
+ "angle": 0,
+ "content": "| # | Instruction |
| 1 | Please evaluate the image's quality comprehensively and provide your insights. Additionally, outline a step-by-step restoration strategy, specifying the sequence of tasks and model choices for each step. The available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement. The available restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 2 | Analyze the quality of the image comprehensively and provide your insights. Furthermore, propose a restoration strategy by detailing each task and model choice sequentially. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement. The restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 3 | Assess the overall quality of the image and provide a detailed evaluation. Then, design a step-by-step restoration process, specifying tasks and model choices. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 4 | Perform a comprehensive evaluation of the image quality and explain your observations. Additionally, develop a step-by-step restoration plan, identifying tasks and model choices. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 5 | Conduct a thorough analysis of the image's quality and provide your insights. Subsequently, create a restoration strategy step by step, specifying the tasks and model choices. The tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 6 | Evaluate the quality of the image comprehensively and outline your findings. Moreover, formulate a sequential restoration plan, detailing tasks and model selections. Available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 7 | Provide a detailed assessment of the image's quality and share your observations. Then, create a restoration strategy in a step-by-step manner, specifying tasks and models. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 8 | Examine the image comprehensively and describe its quality in detail. Additionally, outline a step-wise restoration process, specifying the tasks and models for each stage. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 9 | Review the image quality comprehensively and present your findings. Furthermore, propose a restoration plan step by step, detailing the tasks and models used. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 10 | Perform a detailed evaluation of the image and provide your insights. Then, construct a step-by-step restoration strategy, specifying the sequence of tasks and the corresponding models. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.925,
+ 0.509,
+ 0.937
+ ],
+ "angle": 0,
+ "content": "24"
+ }
+ ],
+ [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.27,
+ 0.123,
+ 0.728,
+ 0.138
+ ],
+ "angle": 0,
+ "content": "Table 9. Responses generated by GPT-4V using the self-instruct strategy [73]"
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.128,
+ 0.149,
+ 0.869,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "| # | Response |
| 1 | The image depicts an outdoor scene viewed through a wet windshield, with water droplets significantly obscuring the view. The evaluated image exhibits severe distortion caused by the water droplets, resulting in a loss of sharpness and clarity. Details of the background, such as the car ahead and surrounding structures, are blurred and difficult to discern. The image also suffers from uneven focus and lacks proper visibility. |
| 2 | This image portrays an outdoor scene as seen through a wet windshield, where water droplets heavily obstruct the view. The distortion caused by these droplets leads to a noticeable reduction in sharpness and clarity. Background elements, like the car in front and nearby structures, appear heavily blurred and challenging to distinguish. The image further suffers from uneven focus, reducing overall visibility. |
| 3 | An outdoor setting is captured in this image, viewed through a wet windshield. The water droplets significantly obscure the scene, causing severe distortion. Sharpness and clarity are greatly reduced, and background features, such as the car ahead and surrounding buildings, are blurred and hard to recognize. Uneven focus further detracts from the image's overall visibility. |
| 4 | This image showcases an outdoor scene as observed through a wet windshield, with water droplets causing significant obstruction. The distortion leads to substantial clarity and sharpness loss. Details in the background, including the car ahead and nearby structures, are blurred and hard to identify. Additionally, the image has uneven focus, which impairs its overall visibility. |
| 5 | The image illustrates an outdoor view observed through a wet windshield, where water droplets obscure much of the scene. The distortion caused by the droplets results in a notable reduction in sharpness and clarity. Background details, such as the car ahead and surrounding structures, appear blurry and indistinct. Uneven focus further degrades the visibility of the image. |
| 6 | In this image, an outdoor setting is depicted as seen through a wet windshield, with water droplets heavily obscuring the view. The distortion from the droplets results in a considerable loss of sharpness and clarity. Background elements, like the car ahead and nearby buildings, are blurred and hard to discern. The uneven focus of the image adds to its overall lack of visibility. |
| 7 | This image shows an outdoor scene viewed through a wet windshield, with water droplets creating significant visual obstruction. The distortion caused by the droplets reduces the sharpness and clarity of the image. Background features, including the car ahead and the surrounding structures, are blurred and difficult to distinguish. The image also suffers from uneven focus, limiting its visibility. |
| 8 | An outdoor view is displayed in this image, captured through a wet windshield with water droplets blocking much of the scene. The distortion caused by the droplets results in reduced clarity and sharpness. Background details, such as the car in front and nearby structures, are blurred and indistinct. Uneven focus further compromises the visibility of the image. |
| 9 | The image depicts an outdoor scene seen through a wet windshield, where water droplets obscure much of the view. This distortion leads to a significant loss of sharpness and clarity. Background elements like the car ahead and nearby structures are blurred and challenging to distinguish. Uneven focus further contributes to the poor visibility of the image. |
| 10 | This image represents an outdoor scene viewed through a wet windshield, with water droplets obscuring the visual field. The distortion caused by the droplets significantly affects sharpness and clarity, making background features like the car ahead and surrounding structures appear blurred and indistinct. The uneven focus further reduces the overall visibility of the image. |
"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.49,
+ 0.925,
+ 0.508,
+ 0.937
+ ],
+ "angle": 0,
+ "content": "25"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/data/2025/2504_04xxx/2504.04158/f2103d47-786d-466a-b93a-76660f3198b5_origin.pdf b/data/2025/2504_04xxx/2504.04158/f2103d47-786d-466a-b93a-76660f3198b5_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..ec5fa5cc0b456b0ae8bd277e8383489387a4f787
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/f2103d47-786d-466a-b93a-76660f3198b5_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7811fcc834fbaa8c6afe4c8075cbc75f12d11ce1617f6fb58d8159d01b68ca01
+size 42136120
diff --git a/data/2025/2504_04xxx/2504.04158/full.md b/data/2025/2504_04xxx/2504.04158/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..ce48dd51727f20797cd26add8b726c7460b18abb
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/full.md
@@ -0,0 +1,654 @@
+# JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration
+
+Yunlong $\mathrm{Lin}^{1*}$ Zixu Lin $^{1*}$ Haoyu Chen $^{2*}$ Panwang Pan $^{3*}$ Chenxin Li $^{6}$
+
+Sixiang Chen² Yeying Jin⁴ Wenbo Li⁵† Xinghao Ding¹
+
+| Task | Tools | Model Description |
| Super-resolution | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Denoising | SCUnet [94] | Hybrid UNet-based model combining convolution and transformer blocks, designed for robust denoising under diverse real-world noise conditions. |
| Compression artifact removal | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Deblurring | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently.. |
| Deraining | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| UDR-S2Former [8] | An uncertainty-aware transformer model for rain streak removal. |
| Img2img-turbo-rain [52] | Efficient model based on SD-turbo, designed for fast and effective rain removal in real-world images. |
| Raindrop removal | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| Dehazing | RIDCP [81] | Efficient dehazing model utilizing high-quality codebook priors to handle complex real-world haze. |
| KANet [21] | Efficient dehazing network using a localization-and-removal pipeline to handle complex real-world hazy. |
| Desnowing | Img2img-turbo-snow [52] | Efficient model for removing snow artifacts while preserving natural scene details. |
| Snowformer [11] | Transformer-based model for removing snowflakes while preserving natural scene details. |
| Low-light enhancement | Img2img-turbo-night [52] | Fast and efficient model based on SD-turbo, designed for low-light enhancement in real-world scenarios. |
| HVI-CIDNet [83] | Lightweight transformer for low-light and exposure correction, enhancing both image quality and downstream vision tasks efficiently. |
| LightenDiff [27] | Diffusion-based framework for low-light enhancement, leveraging Retinex theory and latent-space decomposition for high-quality unsupervised restoration. |
+
+come from diverse sources, including the aforementioned autonomous driving datasets. Additionally, to enhance the generalizability of JarvisIR in natural contexts, we incorporated natural adverse weather scenes from internet and public datasets [5, 34, 39, 42, 47, 57, 87, 99].
+
+# 9.2. Details of degradation library
+
+As described in Sec 3.1 of the manuscript, we simulate realistic adverse weather scenarios—rainy, nighttime, snowy, and foggy conditions—by customizing a degradation library developed with physical models and image transformation techniques to synthesize degraded images. In this section, we detail our degradation implementations, cov
+
+
+Figure 8. Adverse weather scene simulator. To simulate realistic adverse weather scenarios, including rainy, nighttime, snowy, and foggy, we customized the degradation library developed using physical models and image transformation techniques to synthesize degraded images.
+
+ering the principles, formulas, and severity setups for the Night Scene Simulator, Fog Scene Simulator, Rain Scene Simulator, and Snow Scene Simulator. Examples for each implementation are provided in Figure 11.
+
+Night Scene Simulator. Inspired by the work of [19], we employ a low-light degradation transform to synthesize realistic low-light images, denoted as $T_{\text{night}}$ , as illustrated in Figure 8. Specifically, we first convert the daytime image $I_{\text{day}}$ into RAW data using the sRGB $\rightarrow$ RAW process [3]. Next, we linearly attenuate the RAW image and introduce Shot and Read (S&R) noise, which is commonly observed in camera imaging systems [3]. Finally, we apply the Image Signal Processing (ISP) pipeline to convert the low-light sensor data back into sRGB format. Additionally, we incorporate flare degradation using flare templates from the Flare7K++ [20] dataset. The complete low-light degrada
+
+tion transform $T_{\text{night}}$ is given by:
+
+$$
+T _ {\text {n i g h t}} \left(\mathrm {I} _ {\text {d a y}}\right) = T _ {I S P} \left(T _ {s R G B \rightarrow R A W} \left(\mathrm {I} _ {\text {d a y}}\right) + \mathrm {I} _ {\text {n o i s e}}\right) + \mathrm {I} _ {\text {f l a r e}}, \tag {8}
+$$
+
+which generates a degraded image $I_{day}$ that closely resembles a dark nighttime scene. Furthermore, we use an online dynamic degradation process. It applies randomized parameter combinations, as defined in Equation 8, to simulate diverse nighttime driving conditions.
+
+Fog Scene Simulator. Inspired by RIDCP [81], we design a foggy image degradation transform, denoted as $T_{\mathrm{fog}}$ , to synthesize realistic hazy images, as shown in Figure 8. Specifically, we simulate fog by introducing transmission maps $t(x)$ using depth estimation algorithms (e.g., Depth anything V2 [84]), combined with exponential attenuation $e^{\beta d(x)}$ , where $\beta$ controls haze density within the range [0.3, 1.5]. Additionally, poor lighting conditions are
+
+modeled by applying a brightness adjustment factor $\gamma \in [1.5,3.0]$ , Gaussian noise $\mathcal{N}$ , and atmospheric light variation $A + \Delta A$ , where $\Delta A$ is sampled from $[-0.025,0.025]$ . To further enhance realism, JPEG compression artifacts are introduced by applying JPEG $(\cdot)$ to the degraded image. The complete foggy image synthesis process is defined as:
+
+$$
+T _ {\text {f o g}} \left(I _ {\text {d a y}}\right) = \operatorname {J P E G} \left(\mathcal {P} \left(I _ {\text {d a y}} ^ {\gamma} + \mathcal {N}, e ^ {\beta d (x)}, A + \Delta A\right)\right), \tag {9}
+$$
+
+where $\mathcal{P}$ represents the hazy image formation process, $I_{\mathrm{day}}$ is the clean image, and $d(x)$ is the estimated depth map. The variable $x$ refers to the spatial coordinates of the image. This dynamic degradation process is designed to operate online with randomized parameters, simulating diverse real-world foggy conditions.
+
+Rain Scene Simulator. Inspired by PGDGN [54], we introduce a rain degradation transform, denoted as $T_{\mathrm{rain}}$ , to generate realistic rainy images (Figure 8). This transform synthesizes rainy images by combining a disentangled clean image with a physics-based rain rendering model. The degradation process is formulated as:
+
+$$
+T _ {\text {r a i n}} \left(I _ {\text {d a y}}\right) = W _ {\text {M o d}} \left(G \left(I _ {\text {d a y}}\right), \hat {w}, z\right), \tag {10}
+$$
+
+where $I_{\mathrm{day}}$ is the clean image, $G(I_{\mathrm{day}})$ represents the disentangled base image, and $W_{\mathrm{Mod}}$ is the rain rendering model. $W_{\mathrm{Mod}}$ incorporates parameters $\hat{w} = \{\hat{w}_d, \hat{w}_{nd}\}$ , with $\hat{w}_d$ controlling differentiable aspects such as raindrop size and streak density, and $\hat{w}_{nd}$ addressing nondifferentiable properties. The term $z$ introduces stochastic noise for variability in rain effects. This process applies $W_{\mathrm{Mod}}$ to add realistic raindrop occlusions, rain streaks, and scene wetness to the disentangled image. $G(I_{\mathrm{day}})$ , generating a visually plausible rainy image $T_{\mathrm{rain}}(I_{\mathrm{day}})$ with controlled and diverse effects.
+
+Snow Scene Simulator. Building on the img2img-turbo model [52], we introduce a snow transformation, denoted as $T_{\mathrm{snow}}$ , to generate realistic snowy images from daytime inputs. This process uses the SD-Turbo model with textual conditioning. It synthesizes snowy scenes by combining the input image with a latent diffusion-based generator and a textual prompt. The snow transformation is formulated as:
+
+$$
+T _ {\text {s n o w}} \left(I _ {\text {d a y}}, C _ {\text {s n o w}}\right) = G _ {\text {s n o w}} \left(I _ {\text {d a y}}, C _ {\text {s n o w}}\right), \tag {11}
+$$
+
+where $I_{\mathrm{day}}$ is the daytime input image, $C_{\mathrm{snow}}$ is the textual condition (e.g., "driving in the heavy snow"), and $G_{\mathrm{snow}}$ represents the generator. By employing LoRA adapters and skip connections, the generator enables precise control over scene characteristics while maintaining the structural integrity of the input image. This process applies $G_{\mathrm{snow}}$ to infuse the daytime image $I_{\mathrm{day}}$ with snowy features, guided by the contextual information in $C_{\mathrm{snow}}$ . The resulting synthetic image aligns closely with the visual expectations of a snowy environment while maintaining consistency with the original scene's structure.
+
+# 10. More ablation
+
+To thoroughly investigate the proposed JarvisIR, we conducted an extensive array of ablation studies on the CleanBench-Real dataset. Four non-reference metrics are used for assessment: MUSIQ [35], MANIQA [86], CLIP-IQA+ [69], LIQE [95]. The specific elements of these studies are further expounded in the sections that follow.
+
+# 10.1. MRRHF vs. vanilla RRHF
+
+We evaluate the effectiveness of our proposed MRRHF by comparing it with vanilla RRHF [93]. The reward and diversity metrics over training iterations are illustrated in Table 5. Fine-tuning JarvisIR with MRRHF significantly improves the average values of both reward and diversity by 0.19 and 3.43, respectively, compared to using RRHF. The degradation in diversity and reward when using vanilla RRHF results from its offline sample generation strategy. As discussed in Sec. 5 in the manuscript, this strategy confines its generated samples to the finite sample space created by the SFT model using diverse beam search [68]. In contrast, our MRRHF employs a hybrid sample generation strategy and entropy regularization, providing sufficient sample exploration space to achieve globally optimal results.
+
+# 10.2. Sample generation strategy and entropy regularization
+
+In our manuscript, we examine the effects of the sample generation strategy and entropy regularization on the MR-RHF tuning process, focusing on reward scores and response diversity. This section provides further evidence of the effectiveness of our hybrid sample generation strategy and entropy regularization. Specifically, as shown in Table 5, we assess their impact on performance using the CleanBench-Real validation set. The results demonstrate that our hybrid sampling approach and entropy regularization not only enhance training stability and facilitate high-quality exploration of the optimization space but also significantly improve testing performance.
+
+# 10.3. Effectiveness of differentiated contrast weights
+
+In Equation 4 of our manuscript, we refine the original ranking loss [93] by introducing differentiated contrast weights, expressed as $L_{\mathrm{rank}} = \sum_{s_i < s_j} (s_j - s_i) \max(0, p_i - p_j)$ . The term $(s_j - s_i)$ represents the differentiated contrast weights. We compare this with the original ranking loss $\hat{L}_{\mathrm{rank}} = \sum_{s_i < s_j} \max(0, p_i - p_j)$ . Table 5 presents the reward and diversity metrics over training iterations. When the ranking loss is applied without differentiated contrast weights $\hat{L}_{\mathrm{rank}}$ the average values of both reward and diversity decrease by 0.14 and 3.93, respectively, compared
+
+Table 5. Ablation studies on tuning paradigm, differentiated contrast weights, different sample generation strategies, and entropy regularization. The "Reward" represents the average reward scores obtained during alignment tuning, spanning from -1 to 1. A negative score indicates a penalty, while a positive score represents a reward. The "Diversity" reflects the average number of unique responses produced during the training process. Additionally, we evaluate performance on the CleanBench-Real validation set using four non-reference metrics: MUSIQ, MANIQA, CLIP-IQA+, and LIQE. The reported values represent the average performance across all tested scenes.
+
+| # | Instruction |
| 1 | Please evaluate the image's quality comprehensively and provide your insights. Additionally, outline a step-by-step restoration strategy, specifying the sequence of tasks and model choices for each step. The available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement. The available restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 2 | Analyze the quality of the image comprehensively and provide your insights. Furthermore, propose a restoration strategy by detailing each task and model choice sequentially. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement. The restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 3 | Assess the overall quality of the image and provide a detailed evaluation. Then, design a step-by-step restoration process, specifying tasks and model choices. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 4 | Perform a comprehensive evaluation of the image quality and explain your observations. Additionally, develop a step-by-step restoration plan, identifying tasks and model choices. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 5 | Conduct a thorough analysis of the image's quality and provide your insights. Subsequently, create a restoration strategy step by step, specifying the tasks and model choices. The tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 6 | Evaluate the quality of the image comprehensively and outline your findings. Moreover, formulate a sequential restoration plan, detailing tasks and model selections. Available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 7 | Provide a detailed assessment of the image's quality and share your observations. Then, create a restoration strategy in a step-by-step manner, specifying tasks and models. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 8 | Examine the image comprehensively and describe its quality in detail. Additionally, outline a step-wise restoration process, specifying the tasks and models for each stage. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 9 | Review the image quality comprehensively and present your findings. Furthermore, propose a restoration plan step by step, detailing the tasks and models used. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 10 | Perform a detailed evaluation of the image and provide your insights. Then, construct a step-by-step restoration strategy, specifying the sequence of tasks and the corresponding models. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
+
+Table 9. Responses generated by GPT-4V using the self-instruct strategy [73]
+
+| # | Response |
| 1 | The image depicts an outdoor scene viewed through a wet windshield, with water droplets significantly obscuring the view. The evaluated image exhibits severe distortion caused by the water droplets, resulting in a loss of sharpness and clarity. Details of the background, such as the car ahead and surrounding structures, are blurred and difficult to discern. The image also suffers from uneven focus and lacks proper visibility. |
| 2 | This image portrays an outdoor scene as seen through a wet windshield, where water droplets heavily obstruct the view. The distortion caused by these droplets leads to a noticeable reduction in sharpness and clarity. Background elements, like the car in front and nearby structures, appear heavily blurred and challenging to distinguish. The image further suffers from uneven focus, reducing overall visibility. |
| 3 | An outdoor setting is captured in this image, viewed through a wet windshield. The water droplets significantly obscure the scene, causing severe distortion. Sharpness and clarity are greatly reduced, and background features, such as the car ahead and surrounding buildings, are blurred and hard to recognize. Uneven focus further detracts from the image's overall visibility. |
| 4 | This image showcases an outdoor scene as observed through a wet windshield, with water droplets causing significant obstruction. The distortion leads to substantial clarity and sharpness loss. Details in the background, including the car ahead and nearby structures, are blurred and hard to identify. Additionally, the image has uneven focus, which impairs its overall visibility. |
| 5 | The image illustrates an outdoor view observed through a wet windshield, where water droplets obscure much of the scene. The distortion caused by the droplets results in a notable reduction in sharpness and clarity. Background details, such as the car ahead and surrounding structures, appear blurry and indistinct. Uneven focus further degrades the visibility of the image. |
| 6 | In this image, an outdoor setting is depicted as seen through a wet windshield, with water droplets heavily obscuring the view. The distortion from the droplets results in a considerable loss of sharpness and clarity. Background elements, like the car ahead and nearby buildings, are blurred and hard to discern. The uneven focus of the image adds to its overall lack of visibility. |
| 7 | This image shows an outdoor scene viewed through a wet windshield, with water droplets creating significant visual obstruction. The distortion caused by the droplets reduces the sharpness and clarity of the image. Background features, including the car ahead and the surrounding structures, are blurred and difficult to distinguish. The image also suffers from uneven focus, limiting its visibility. |
| 8 | An outdoor view is displayed in this image, captured through a wet windshield with water droplets blocking much of the scene. The distortion caused by the droplets results in reduced clarity and sharpness. Background details, such as the car in front and nearby structures, are blurred and indistinct. Uneven focus further compromises the visibility of the image. |
| 9 | The image depicts an outdoor scene seen through a wet windshield, where water droplets obscure much of the view. This distortion leads to a significant loss of sharpness and clarity. Background elements like the car ahead and nearby structures are blurred and challenging to distinguish. Uneven focus further contributes to the poor visibility of the image. |
| 10 | This image represents an outdoor scene viewed through a wet windshield, with water droplets obscuring the visual field. The distortion caused by the droplets significantly affects sharpness and clarity, making background features like the car ahead and surrounding structures appear blurred and indistinct. The uneven focus further reduces the overall visibility of the image. |
\ No newline at end of file
diff --git a/data/2025/2504_04xxx/2504.04158/images/0053739e7b4a0daff36d611a58d57cf49f0f83e2bc195074da33afa7f82a02c2.jpg b/data/2025/2504_04xxx/2504.04158/images/0053739e7b4a0daff36d611a58d57cf49f0f83e2bc195074da33afa7f82a02c2.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..674b34988e7ff93647ebdd3e5750bec5a6b5c164
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/0053739e7b4a0daff36d611a58d57cf49f0f83e2bc195074da33afa7f82a02c2.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:729a30883f76b4b670dc1fa0ca4300f9aaccdd01955cfc7f7e75c1f423e6b16b
+size 4408
diff --git a/data/2025/2504_04xxx/2504.04158/images/048f68f8d200027e9e905194273c996241437124d691fadea0d3323580d729c0.jpg b/data/2025/2504_04xxx/2504.04158/images/048f68f8d200027e9e905194273c996241437124d691fadea0d3323580d729c0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..65e6bcd3dcd386a1cae5471490c0f83e7278296e
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/048f68f8d200027e9e905194273c996241437124d691fadea0d3323580d729c0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32ded931743642b6f0fcdd7aa1af58078de51e5316c3eb99d12bdc77621ef796
+size 99836
diff --git a/data/2025/2504_04xxx/2504.04158/images/07e47185cabd4f6aa069f0731377374d3c4c4a1803e330eba79cb41be97b828b.jpg b/data/2025/2504_04xxx/2504.04158/images/07e47185cabd4f6aa069f0731377374d3c4c4a1803e330eba79cb41be97b828b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..bff21c6320d4609a84a769c1f189168d7fabfebf
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/07e47185cabd4f6aa069f0731377374d3c4c4a1803e330eba79cb41be97b828b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18fecada100f934c9f5e662219de76e0f337229541606c7dc4f6a2c9b032a6d6
+size 21105
diff --git a/data/2025/2504_04xxx/2504.04158/images/0a3b5b7d535c5b7854fd3bdde961cf5b698ef5e7f73461df68da29efbd4ab814.jpg b/data/2025/2504_04xxx/2504.04158/images/0a3b5b7d535c5b7854fd3bdde961cf5b698ef5e7f73461df68da29efbd4ab814.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e35ffe3afabce365e4c3eb3cfdbab0002c00760e
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/0a3b5b7d535c5b7854fd3bdde961cf5b698ef5e7f73461df68da29efbd4ab814.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:037587af77385f4cde1607295c2fd92ef643894366dbf39331524ba2799764b4
+size 4443
diff --git a/data/2025/2504_04xxx/2504.04158/images/0e5ce300908b57f9feceb059f8e972fcafc8937282b92fdcf6eeb2742dfefbff.jpg b/data/2025/2504_04xxx/2504.04158/images/0e5ce300908b57f9feceb059f8e972fcafc8937282b92fdcf6eeb2742dfefbff.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..901847630c6fe9b0f2bf52a4d55641997c23cfe1
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/0e5ce300908b57f9feceb059f8e972fcafc8937282b92fdcf6eeb2742dfefbff.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bb3687db0b4a5c36491a4459c4292dcf94aa3686f2228219000bee0b0aa0fa5
+size 5353
diff --git a/data/2025/2504_04xxx/2504.04158/images/0f2a9776b0ac32e01010f99a1c6630981a7872f316a9fdeaa8ce9b3157842225.jpg b/data/2025/2504_04xxx/2504.04158/images/0f2a9776b0ac32e01010f99a1c6630981a7872f316a9fdeaa8ce9b3157842225.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3e1f7f3d533142d47b355d69f32c370fd26b072b
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/0f2a9776b0ac32e01010f99a1c6630981a7872f316a9fdeaa8ce9b3157842225.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e1faf84bb972cd9d61f41f11cd9a37caf915a9d7bb7205831940c07ec8290b5
+size 4413
diff --git a/data/2025/2504_04xxx/2504.04158/images/0fee2318775c06e33b9fc28cd1255ba845f9a9a426d4ef15368be65a8ecc85be.jpg b/data/2025/2504_04xxx/2504.04158/images/0fee2318775c06e33b9fc28cd1255ba845f9a9a426d4ef15368be65a8ecc85be.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2ce71a1dd86b4e20a5b410d2697c666ce97179f3
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/0fee2318775c06e33b9fc28cd1255ba845f9a9a426d4ef15368be65a8ecc85be.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:534150d352f819295463c3f56bb4aaf0ee123cfdd092d742511df64251a40d53
+size 5048
diff --git a/data/2025/2504_04xxx/2504.04158/images/1092add7be2d629c4e97c409e32e011770039f37010cfecbcfd5b249e210e764.jpg b/data/2025/2504_04xxx/2504.04158/images/1092add7be2d629c4e97c409e32e011770039f37010cfecbcfd5b249e210e764.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d9b3daa1e2c08b8f34d9e935a962d1cd405e3815
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/1092add7be2d629c4e97c409e32e011770039f37010cfecbcfd5b249e210e764.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70e354763e96a754d22dcab451f16bad76b41dd8828fdd72dc679e9a03c4233d
+size 5954
diff --git a/data/2025/2504_04xxx/2504.04158/images/196e21b799360885c80ca582ee0b9e4ca7695c3d72ec4c24a1b3ba05b28036d9.jpg b/data/2025/2504_04xxx/2504.04158/images/196e21b799360885c80ca582ee0b9e4ca7695c3d72ec4c24a1b3ba05b28036d9.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9fa4ddf94e70cea7639c3cee22d833a75cbf102f
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/196e21b799360885c80ca582ee0b9e4ca7695c3d72ec4c24a1b3ba05b28036d9.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82892302bfcd67f9d30a5ed9e48ad5d2b32741ccc961c5a7f2ba5f61280cd9a1
+size 174798
diff --git a/data/2025/2504_04xxx/2504.04158/images/207bb169b76aa06ce93fe74912b803baad17e211a9f516731f63e51dc61f3f45.jpg b/data/2025/2504_04xxx/2504.04158/images/207bb169b76aa06ce93fe74912b803baad17e211a9f516731f63e51dc61f3f45.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..78ce832d53d0a769e42f749913839fa15f92d4d4
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/207bb169b76aa06ce93fe74912b803baad17e211a9f516731f63e51dc61f3f45.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d962e3c334013a400b04acf6605421a9c19bef86b8388bce737604d75a89bd5b
+size 1535
diff --git a/data/2025/2504_04xxx/2504.04158/images/277db720d3cce45dca68c37a6576170fd2c9bf4bd23cdd7ed198441f6cc70472.jpg b/data/2025/2504_04xxx/2504.04158/images/277db720d3cce45dca68c37a6576170fd2c9bf4bd23cdd7ed198441f6cc70472.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..aeb59f576282b43bf9807969a8b1b9ffd94f9240
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/277db720d3cce45dca68c37a6576170fd2c9bf4bd23cdd7ed198441f6cc70472.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14ff31d4346568ddb364896fb220f90222bd23c3f015519531b33230120f1d92
+size 22728
diff --git a/data/2025/2504_04xxx/2504.04158/images/2782c78e07bdd618db0f57c0df1cbc2cedfb60ec2c542a163c746b4b6c60921e.jpg b/data/2025/2504_04xxx/2504.04158/images/2782c78e07bdd618db0f57c0df1cbc2cedfb60ec2c542a163c746b4b6c60921e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..914463c62980c31cebf56aa663ae67daf62a030f
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/2782c78e07bdd618db0f57c0df1cbc2cedfb60ec2c542a163c746b4b6c60921e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7dd8375bd895ba0081410411166b92286c4142b6bbd60b0a7da876179b05d25
+size 5208
diff --git a/data/2025/2504_04xxx/2504.04158/images/27dbe72546925137215698d71342aa4b9a852bafee62032063e269c7cd2e74e4.jpg b/data/2025/2504_04xxx/2504.04158/images/27dbe72546925137215698d71342aa4b9a852bafee62032063e269c7cd2e74e4.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fee9c0e6ec89434b9cef4627445272aa8e4f119b
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/27dbe72546925137215698d71342aa4b9a852bafee62032063e269c7cd2e74e4.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66d727f9f5be671cf8b11d74b1dd987e563890103b419c843cf367de2c6d9c48
+size 4695
diff --git a/data/2025/2504_04xxx/2504.04158/images/28ca1c0dc94e2e31660a328313235104158535336b5c3a75df2ca4dfd14a7f96.jpg b/data/2025/2504_04xxx/2504.04158/images/28ca1c0dc94e2e31660a328313235104158535336b5c3a75df2ca4dfd14a7f96.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4031cab367d5e6bc196a3a7102d138d982703ce3
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/28ca1c0dc94e2e31660a328313235104158535336b5c3a75df2ca4dfd14a7f96.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb642b2d3957e07de42887006f82108500e182d8cfbacd71d190f3edba28f028
+size 4661
diff --git a/data/2025/2504_04xxx/2504.04158/images/28dcbcdb856150cde102df404c08742064c42a5fe199c8dad17ae790624edf5e.jpg b/data/2025/2504_04xxx/2504.04158/images/28dcbcdb856150cde102df404c08742064c42a5fe199c8dad17ae790624edf5e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..49dfd8a6304f97d0e348d609dc59eb39e0ff44f5
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/28dcbcdb856150cde102df404c08742064c42a5fe199c8dad17ae790624edf5e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:311cf9bc45792df34a274c8b21f6ff89676583014ae9d0426ce4223b16a3fe40
+size 6283
diff --git a/data/2025/2504_04xxx/2504.04158/images/2a22d08258bf1259e15883edf9f8f9154005b5159041b01ecb78994307b1f5cf.jpg b/data/2025/2504_04xxx/2504.04158/images/2a22d08258bf1259e15883edf9f8f9154005b5159041b01ecb78994307b1f5cf.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..0e6bbd9e1de5ef1ad80eef1d23fe140a9881485e
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/2a22d08258bf1259e15883edf9f8f9154005b5159041b01ecb78994307b1f5cf.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d32a7faff9f4e818643672ceab6c215006856e57e35123990bd1f9e0f199d2d8
+size 9634
diff --git a/data/2025/2504_04xxx/2504.04158/images/3167daa672b1ba80ecfa48565a3bde6433ca52b433260e98a606ab32a072f17a.jpg b/data/2025/2504_04xxx/2504.04158/images/3167daa672b1ba80ecfa48565a3bde6433ca52b433260e98a606ab32a072f17a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3eae589cbf39929ed78b77c5c585c56b1e29fccf
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/3167daa672b1ba80ecfa48565a3bde6433ca52b433260e98a606ab32a072f17a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc09ac5853d19756d08330adb0f3217ff0431f2f29bd810ab452470ed2cb935d
+size 11887
diff --git a/data/2025/2504_04xxx/2504.04158/images/36f3a9b8917dbdaac4d3d9ee738ef89e9b7f563ea50eb80cfe269088ee435d04.jpg b/data/2025/2504_04xxx/2504.04158/images/36f3a9b8917dbdaac4d3d9ee738ef89e9b7f563ea50eb80cfe269088ee435d04.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..1e57a05519561976d92877c1181aecf094275f2b
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/36f3a9b8917dbdaac4d3d9ee738ef89e9b7f563ea50eb80cfe269088ee435d04.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:573313b6274153d0f82131de619e4f6ab937cb8ffeb0bb8b818fc342d6bb2ca5
+size 5012
diff --git a/data/2025/2504_04xxx/2504.04158/images/3a64ac6bc4ec1914f34d6d96ae7b3761a8bedd024622ff6c865c177079eb1243.jpg b/data/2025/2504_04xxx/2504.04158/images/3a64ac6bc4ec1914f34d6d96ae7b3761a8bedd024622ff6c865c177079eb1243.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d2db00ec1c39df24a46d39a3c2bf42d7be8b5498
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/3a64ac6bc4ec1914f34d6d96ae7b3761a8bedd024622ff6c865c177079eb1243.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a9ac8ce3e78c2687f533e09f4261cd3c8fa436701083d85ad8d9dd41d0b2350
+size 4396
diff --git a/data/2025/2504_04xxx/2504.04158/images/3b1b2f8721129799a01f82887d8cd7af6101c2959bfad4999bac7002aa393cb7.jpg b/data/2025/2504_04xxx/2504.04158/images/3b1b2f8721129799a01f82887d8cd7af6101c2959bfad4999bac7002aa393cb7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..57b22672e8b18bf47de5220a3614852657a85dcb
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/3b1b2f8721129799a01f82887d8cd7af6101c2959bfad4999bac7002aa393cb7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7838e4be77e8a7f78c7e819bec7a4d749497651d4301ea14e75a23227844c776
+size 8021
diff --git a/data/2025/2504_04xxx/2504.04158/images/3ce7947a2594a77b7e6229dd1b8f04d4a9a1b99c63bb6ba190b22f88bd9ea21c.jpg b/data/2025/2504_04xxx/2504.04158/images/3ce7947a2594a77b7e6229dd1b8f04d4a9a1b99c63bb6ba190b22f88bd9ea21c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..7e96e15784a064d0f2dd771219003bfdf9a1f1b8
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/3ce7947a2594a77b7e6229dd1b8f04d4a9a1b99c63bb6ba190b22f88bd9ea21c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0da38ee379f7225fe5447b8554212685e986968a84364ca6859645a4e374e00
+size 7692
diff --git a/data/2025/2504_04xxx/2504.04158/images/46e0fd7cfb79df6a4f73542600a61049ba8364fa75e365b59aaf3d2e9e3a372c.jpg b/data/2025/2504_04xxx/2504.04158/images/46e0fd7cfb79df6a4f73542600a61049ba8364fa75e365b59aaf3d2e9e3a372c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3153351162e9ccda39139b46db838fc78b86a506
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/46e0fd7cfb79df6a4f73542600a61049ba8364fa75e365b59aaf3d2e9e3a372c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89896280e68996900ad6692515f1056c0a80b8aeffcb6af4df4cfe3f7921d0bb
+size 473212
diff --git a/data/2025/2504_04xxx/2504.04158/images/47e51147399e5c5cac4d772fd2b947ea7e0a56f44388eb2bf47573be6d363c73.jpg b/data/2025/2504_04xxx/2504.04158/images/47e51147399e5c5cac4d772fd2b947ea7e0a56f44388eb2bf47573be6d363c73.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..dbfde742d547885c167f30c819670c781f170655
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/47e51147399e5c5cac4d772fd2b947ea7e0a56f44388eb2bf47573be6d363c73.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:223ae9339a07e134fa3d22f8046356e530ec7d72b5672e5041c390c141ab3803
+size 31958
diff --git a/data/2025/2504_04xxx/2504.04158/images/49f44d4c8efe33a28fa0e42a3441429805f47b0b0e70ed4f4966e19fe9beddce.jpg b/data/2025/2504_04xxx/2504.04158/images/49f44d4c8efe33a28fa0e42a3441429805f47b0b0e70ed4f4966e19fe9beddce.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..78549b837d1868a8755ce0ee45cc1aeda079deb0
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/49f44d4c8efe33a28fa0e42a3441429805f47b0b0e70ed4f4966e19fe9beddce.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2e2b6f7c1119420ac001ff873edb56810f31189c79c6e0db75e54bff35348e3
+size 8096
diff --git a/data/2025/2504_04xxx/2504.04158/images/49f864eff44e607249aefe483cf3a1037af9f29bdaeb255702bdf9f6627aedc4.jpg b/data/2025/2504_04xxx/2504.04158/images/49f864eff44e607249aefe483cf3a1037af9f29bdaeb255702bdf9f6627aedc4.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c274b5932ddd096fb06fd277d40f23e5e25730fe
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/49f864eff44e607249aefe483cf3a1037af9f29bdaeb255702bdf9f6627aedc4.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ece4220711a07d61b460ecd1544fe903275299156a092856252207a184e8566f
+size 1238
diff --git a/data/2025/2504_04xxx/2504.04158/images/597c14e6ec4a37e1a335b094f922b79629803c2edf1b1f6c77e44916153b1f97.jpg b/data/2025/2504_04xxx/2504.04158/images/597c14e6ec4a37e1a335b094f922b79629803c2edf1b1f6c77e44916153b1f97.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..832870a278fc135b40b9ef1efa1420b61ef0210b
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/597c14e6ec4a37e1a335b094f922b79629803c2edf1b1f6c77e44916153b1f97.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:817abc44d573ecc14c72ef37e8f02e627f265c738aa806e98318f7e0e56d80ae
+size 30687
diff --git a/data/2025/2504_04xxx/2504.04158/images/5d8d0415d57fd704b5777e7635db1b3be3e14955cd8984f32baac1558eb4fddb.jpg b/data/2025/2504_04xxx/2504.04158/images/5d8d0415d57fd704b5777e7635db1b3be3e14955cd8984f32baac1558eb4fddb.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9f42598e58f3bc4249659da74c2c09e3b6855f56
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/5d8d0415d57fd704b5777e7635db1b3be3e14955cd8984f32baac1558eb4fddb.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:699d1e98474238efccd1ff5c15b4e30e3c5fa0222915a3900199baa0464517a3
+size 4391
diff --git a/data/2025/2504_04xxx/2504.04158/images/5daf42f90dfb4eed1d4e72358804e3758b194cd830e8d45020c22b208186c24c.jpg b/data/2025/2504_04xxx/2504.04158/images/5daf42f90dfb4eed1d4e72358804e3758b194cd830e8d45020c22b208186c24c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ccdc0603891a01e4104d9c0c8c3e8ba99a6d4850
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/5daf42f90dfb4eed1d4e72358804e3758b194cd830e8d45020c22b208186c24c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdd55e8433ce629c07a0e08155865f95bf021f832033d665c57c96c6a4156c77
+size 54219
diff --git a/data/2025/2504_04xxx/2504.04158/images/5fae250c32fe9d83100696bcfd757e6c5a3418f6108650c849ad9dcd9364d18b.jpg b/data/2025/2504_04xxx/2504.04158/images/5fae250c32fe9d83100696bcfd757e6c5a3418f6108650c849ad9dcd9364d18b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..fed37415fe83e45bc3284d1b4de9d00763e816c7
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/5fae250c32fe9d83100696bcfd757e6c5a3418f6108650c849ad9dcd9364d18b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a542c0d1c8f0817e48a6b73c4b453c66e082bbc1ecb12a9316b7a4faa90a3334
+size 9866
diff --git a/data/2025/2504_04xxx/2504.04158/images/6021787e5d67446e42eb8db90760ca4675db5019ebeb4d140a1597b4575bc329.jpg b/data/2025/2504_04xxx/2504.04158/images/6021787e5d67446e42eb8db90760ca4675db5019ebeb4d140a1597b4575bc329.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a9a6602168bedecb084d6173ca0df5ff39d1147d
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/6021787e5d67446e42eb8db90760ca4675db5019ebeb4d140a1597b4575bc329.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f3f9275a55292c2352243e5a17a96869aafd88903a6681a773010b864f5f410
+size 4033
diff --git a/data/2025/2504_04xxx/2504.04158/images/60462e12727c65b334568dd0fcad44d4540154741c5ebf2eacdb709fcfd0825c.jpg b/data/2025/2504_04xxx/2504.04158/images/60462e12727c65b334568dd0fcad44d4540154741c5ebf2eacdb709fcfd0825c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2e1673ffb4b89cfaec51162ebc92257e01d18516
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/60462e12727c65b334568dd0fcad44d4540154741c5ebf2eacdb709fcfd0825c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31aa62464182b0166a97aa2cf923459cce35cf22245d4fa387c8f81ec44821b2
+size 4119
diff --git a/data/2025/2504_04xxx/2504.04158/images/630d636b5d4242f1d150ff18dc658c28152765add4f18181e0b58f025df8076f.jpg b/data/2025/2504_04xxx/2504.04158/images/630d636b5d4242f1d150ff18dc658c28152765add4f18181e0b58f025df8076f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4cd0aec25a431b6f9afce2ceb2776d606fce5907
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/630d636b5d4242f1d150ff18dc658c28152765add4f18181e0b58f025df8076f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc0e733bfbf013d0f5592c99fe3d00bd07355e5dc2d7641a4b3aa29a800a45c9
+size 4282
diff --git a/data/2025/2504_04xxx/2504.04158/images/63299634de3ab81c0cf9b53336c2320fef84a49cd01c9e26be6e0cc5f371dae1.jpg b/data/2025/2504_04xxx/2504.04158/images/63299634de3ab81c0cf9b53336c2320fef84a49cd01c9e26be6e0cc5f371dae1.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..8858fab6dd3f2da688d4fe86b5efe8911a503c8e
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/63299634de3ab81c0cf9b53336c2320fef84a49cd01c9e26be6e0cc5f371dae1.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e7ddcc885b45d4846ef15a1e2660ef0809b4565b248827744fd49592c023526
+size 201979
diff --git a/data/2025/2504_04xxx/2504.04158/images/69e7782109021642b8ad0d85256034142eddded932f3031d8ce6fd85cf78faa3.jpg b/data/2025/2504_04xxx/2504.04158/images/69e7782109021642b8ad0d85256034142eddded932f3031d8ce6fd85cf78faa3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..48779bb9645868defe0dc65c62aa932b73981937
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/69e7782109021642b8ad0d85256034142eddded932f3031d8ce6fd85cf78faa3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:117503c8a20e220db6c4c17add0755274d46a785bce927dd506a5e64ed20d595
+size 129563
diff --git a/data/2025/2504_04xxx/2504.04158/images/6e0db1833b93ede94fc0b2b3fb4f6bf77e08c72f80ae29c558897232d25b3fb2.jpg b/data/2025/2504_04xxx/2504.04158/images/6e0db1833b93ede94fc0b2b3fb4f6bf77e08c72f80ae29c558897232d25b3fb2.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..6fef8295a400f202e1a248a6264fcf02c7c56a0c
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/6e0db1833b93ede94fc0b2b3fb4f6bf77e08c72f80ae29c558897232d25b3fb2.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c77656bceec480d063dffec915bea53207c70308f7bbac26e1529c247046f06a
+size 6421
diff --git a/data/2025/2504_04xxx/2504.04158/images/6edc4ec7d8dcb99bd5bd5ee9c18053bbc624d9831e4b1d9524139ad66d9d8acf.jpg b/data/2025/2504_04xxx/2504.04158/images/6edc4ec7d8dcb99bd5bd5ee9c18053bbc624d9831e4b1d9524139ad66d9d8acf.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..97359f5a197a383d4ac8087aaec5ddf0b01ee8df
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/6edc4ec7d8dcb99bd5bd5ee9c18053bbc624d9831e4b1d9524139ad66d9d8acf.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5258dc9db2ee08a777b61c7b5641d244880153c8d84077d3641e1ca926e3202b
+size 6961
diff --git a/data/2025/2504_04xxx/2504.04158/images/71a63f46812d2f4fe58ba8bc6a43938249dabe1c45d6bd32c8ceaa33a43536fc.jpg b/data/2025/2504_04xxx/2504.04158/images/71a63f46812d2f4fe58ba8bc6a43938249dabe1c45d6bd32c8ceaa33a43536fc.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..31201fa7ebc462c6f6c0d8a74c6086c918c811a6
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/71a63f46812d2f4fe58ba8bc6a43938249dabe1c45d6bd32c8ceaa33a43536fc.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afc437b6408c487b96e73a6b82343f481408444330118c9d9998f1da1a6a0902
+size 172157
diff --git a/data/2025/2504_04xxx/2504.04158/images/7271b51303fa9e1d73ac5a18d3905c3170155aeba56af009421e927c7c322506.jpg b/data/2025/2504_04xxx/2504.04158/images/7271b51303fa9e1d73ac5a18d3905c3170155aeba56af009421e927c7c322506.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..58e2ff8257c969d2a6913e2380456567f3ffa25c
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/7271b51303fa9e1d73ac5a18d3905c3170155aeba56af009421e927c7c322506.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f24678b4f266e275f400cbd537562b7457b638cb279298071146c84d77d62782
+size 365621
diff --git a/data/2025/2504_04xxx/2504.04158/images/755073a76e638737e44f51b897ed3b2159e88baa6bb4c4c892c9d48bb26ede25.jpg b/data/2025/2504_04xxx/2504.04158/images/755073a76e638737e44f51b897ed3b2159e88baa6bb4c4c892c9d48bb26ede25.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d1324c249a62153d0d0f921bc5e1cf740b1389df
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/755073a76e638737e44f51b897ed3b2159e88baa6bb4c4c892c9d48bb26ede25.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fa582d31efaa0fb05c7ece5d9d125ec657baf2e7e1b80c1f99cc5efb917bc66
+size 10524
diff --git a/data/2025/2504_04xxx/2504.04158/images/77e20f9404e45b0bc0c37fa8923a96bbd56bc7f30678dd510041fc41e98e8c44.jpg b/data/2025/2504_04xxx/2504.04158/images/77e20f9404e45b0bc0c37fa8923a96bbd56bc7f30678dd510041fc41e98e8c44.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..72f4e73c24db1df54a727dcbcb616a7809b08f2d
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/77e20f9404e45b0bc0c37fa8923a96bbd56bc7f30678dd510041fc41e98e8c44.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f470f0f6ee0c64cef05b09a2f98469625fd606f91d536fac8343eb5c7c421518
+size 3260
diff --git a/data/2025/2504_04xxx/2504.04158/images/79f7595ecedc932e74b8dd4c53cb640edad0c8959a0c6a35f26230e212642d85.jpg b/data/2025/2504_04xxx/2504.04158/images/79f7595ecedc932e74b8dd4c53cb640edad0c8959a0c6a35f26230e212642d85.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..61dd7be0ff8a429ad79a43fca1feb3d2e7fcd1be
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/79f7595ecedc932e74b8dd4c53cb640edad0c8959a0c6a35f26230e212642d85.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b6a939d5c730556ae684cbccd5ead2b39224335df0e1b38b46348a463078317
+size 40328
diff --git a/data/2025/2504_04xxx/2504.04158/images/7e2e7bd30d3e24c638b2e1e4b83c9f966daefc13553215b4ee179af086047012.jpg b/data/2025/2504_04xxx/2504.04158/images/7e2e7bd30d3e24c638b2e1e4b83c9f966daefc13553215b4ee179af086047012.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2815227b351be21b417fa1a6eb10bb567af630a4
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/7e2e7bd30d3e24c638b2e1e4b83c9f966daefc13553215b4ee179af086047012.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40b6702104ecfe6ebfc9a0efc2ce275182ad656bee365d7dc82bb78d1f9ecb39
+size 4461
diff --git a/data/2025/2504_04xxx/2504.04158/images/891a2dcf761e81f0b04879d2e733b5876e84cffd9a1a36b3d9a4caea1fdc7da6.jpg b/data/2025/2504_04xxx/2504.04158/images/891a2dcf761e81f0b04879d2e733b5876e84cffd9a1a36b3d9a4caea1fdc7da6.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..4a6863e08c6d0c622cae10c8bf0e0c8dc405cda3
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/891a2dcf761e81f0b04879d2e733b5876e84cffd9a1a36b3d9a4caea1fdc7da6.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf696ddaa30d8f7cddc45dd066bd195001fa14593f1a0f5316eb87d9c6bfca00
+size 28572
diff --git a/data/2025/2504_04xxx/2504.04158/images/8a82da58f19f001a460478c5aa34cd0fa33b0a779221ccf559b57ee7b38a3e7f.jpg b/data/2025/2504_04xxx/2504.04158/images/8a82da58f19f001a460478c5aa34cd0fa33b0a779221ccf559b57ee7b38a3e7f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c13f579c8f99fcddfb74ff167c36574cf6abc035
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/8a82da58f19f001a460478c5aa34cd0fa33b0a779221ccf559b57ee7b38a3e7f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdcfca8a3e640ece8e20e59a8941c29f515c4e9476083612dea98f779538fee7
+size 41649
diff --git a/data/2025/2504_04xxx/2504.04158/images/917405f323bfed15e07d97060629637b7c1a50fd3d1629b96986c5c02ab0c32b.jpg b/data/2025/2504_04xxx/2504.04158/images/917405f323bfed15e07d97060629637b7c1a50fd3d1629b96986c5c02ab0c32b.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..09e3adee7a506123a7ec7943d9629686b0fed1e1
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/917405f323bfed15e07d97060629637b7c1a50fd3d1629b96986c5c02ab0c32b.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07f92a61efad083e9ccf8b09dc9a3d25e0186581f538c60a4268e7fdeb2b3d50
+size 5156
diff --git a/data/2025/2504_04xxx/2504.04158/images/989095191339c4f747a6b323daf2680729f045b796b53a1881f9a8bf8e31c769.jpg b/data/2025/2504_04xxx/2504.04158/images/989095191339c4f747a6b323daf2680729f045b796b53a1881f9a8bf8e31c769.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..894a17e85a0ee27e766818205d01b0514bd5de0f
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/989095191339c4f747a6b323daf2680729f045b796b53a1881f9a8bf8e31c769.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1b0246756c878496fcc589acf78047817be024a8277cf58ebcd1040d4969704
+size 7033
diff --git a/data/2025/2504_04xxx/2504.04158/images/99189216912451c3188de984a94af5b49db15f7a32fd2fa5f1020a07dbd19c15.jpg b/data/2025/2504_04xxx/2504.04158/images/99189216912451c3188de984a94af5b49db15f7a32fd2fa5f1020a07dbd19c15.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..de51366c38e73b882b25082525749f449af895db
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/99189216912451c3188de984a94af5b49db15f7a32fd2fa5f1020a07dbd19c15.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc75fa7dbde9a620fae1567a5dd269a8af9a069f9339bd8d62ed7e1f9e03d416
+size 1514
diff --git a/data/2025/2504_04xxx/2504.04158/images/a5c8902fafce1da343f0eebbc19d491c688868f6f7fd94729f9255756a759bd7.jpg b/data/2025/2504_04xxx/2504.04158/images/a5c8902fafce1da343f0eebbc19d491c688868f6f7fd94729f9255756a759bd7.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..b7af9a1b54afd28f62058c0ec84790bd129e4661
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/a5c8902fafce1da343f0eebbc19d491c688868f6f7fd94729f9255756a759bd7.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34da41a5839f7ec6932c5ca1903989d3dfd56aca6fe717fb3bd358a6eda62ce2
+size 7620
diff --git a/data/2025/2504_04xxx/2504.04158/images/a9e951bc184e32175b8255774a3a8ea4d6923dcf0336b8f1c3a9aca75ea669a0.jpg b/data/2025/2504_04xxx/2504.04158/images/a9e951bc184e32175b8255774a3a8ea4d6923dcf0336b8f1c3a9aca75ea669a0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..129c90afc5d27dbcc9aafc46233107b64f7927f3
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/a9e951bc184e32175b8255774a3a8ea4d6923dcf0336b8f1c3a9aca75ea669a0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:486ca45e0f2931cd92e918c105cfc2c04d35611170e5ab954a456b96420f9fb4
+size 69835
diff --git a/data/2025/2504_04xxx/2504.04158/images/aafd2cf00fe6a8d7048b14de8ff253226fcdeda34e9179f2ccb3df4cc9858a94.jpg b/data/2025/2504_04xxx/2504.04158/images/aafd2cf00fe6a8d7048b14de8ff253226fcdeda34e9179f2ccb3df4cc9858a94.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..677483452d2b71ca56241042dc412d2e6020dc28
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/aafd2cf00fe6a8d7048b14de8ff253226fcdeda34e9179f2ccb3df4cc9858a94.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b0ccb6c471fbece9ac4e03ba83bd180f4e526258d958bbc5b12eff7e33f23ef
+size 34449
diff --git a/data/2025/2504_04xxx/2504.04158/images/ad06fd35d424cd4690e4b04bc67fef6cac8c79888a7354a90db33171058c65eb.jpg b/data/2025/2504_04xxx/2504.04158/images/ad06fd35d424cd4690e4b04bc67fef6cac8c79888a7354a90db33171058c65eb.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9270530531414cef03d7040d4e3a766e8a667733
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/ad06fd35d424cd4690e4b04bc67fef6cac8c79888a7354a90db33171058c65eb.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46c88d7d9f6aaf0d68fcf5c2db89cdbd086cc356a76b41cfac5cd1b33232b202
+size 1605
diff --git a/data/2025/2504_04xxx/2504.04158/images/b1da7d6ee75fee18988d32322f285d514a220d35e51b55a14df02fbbbb9b32f0.jpg b/data/2025/2504_04xxx/2504.04158/images/b1da7d6ee75fee18988d32322f285d514a220d35e51b55a14df02fbbbb9b32f0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..e7296e195eac2bfc5280da58f9da9015782cad2a
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/b1da7d6ee75fee18988d32322f285d514a220d35e51b55a14df02fbbbb9b32f0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43bf5f6fa7486beaa675a4ceb2032c676bd8036975c4af778bc95cd8fffa44a5
+size 4157
diff --git a/data/2025/2504_04xxx/2504.04158/images/b5c4a022d3b485a86415dca329195c27b3d76253d0bf5302b288d80ed485bc0a.jpg b/data/2025/2504_04xxx/2504.04158/images/b5c4a022d3b485a86415dca329195c27b3d76253d0bf5302b288d80ed485bc0a.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..53d5f96b25346279c7a27b3e7d419493e66a8d42
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/b5c4a022d3b485a86415dca329195c27b3d76253d0bf5302b288d80ed485bc0a.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85beb9d32a560aee7ff3c9d385578a47a00482c55d15b2064bae80ff1a82d97b
+size 6402
diff --git a/data/2025/2504_04xxx/2504.04158/images/b852dd7f672c01f186f654d79ea9b14b15c39241a7bf2273973581801987640c.jpg b/data/2025/2504_04xxx/2504.04158/images/b852dd7f672c01f186f654d79ea9b14b15c39241a7bf2273973581801987640c.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..960dc48c645d0588ad322f81e34bc84d64bb9c29
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/b852dd7f672c01f186f654d79ea9b14b15c39241a7bf2273973581801987640c.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:188d8cb731d7b95c7f1e2950187a8373ff20e0593e7483cd477d9e8891c0c0a9
+size 4150
diff --git a/data/2025/2504_04xxx/2504.04158/images/bbe9edc792c73d752d01aa5189954bc8a765a5488fe84d4037afc99812d38775.jpg b/data/2025/2504_04xxx/2504.04158/images/bbe9edc792c73d752d01aa5189954bc8a765a5488fe84d4037afc99812d38775.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..b7157475e1b079c7398eefeb9a9229d348c7ca7a
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/bbe9edc792c73d752d01aa5189954bc8a765a5488fe84d4037afc99812d38775.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad545a928def652f5b3ae5615e4a0481369e5c5967c94b820174d25fcdc55408
+size 33388
diff --git a/data/2025/2504_04xxx/2504.04158/images/c4853ff713766c35c51fbc2e430ef6956f66c6a2614f43f85800c319e7f72096.jpg b/data/2025/2504_04xxx/2504.04158/images/c4853ff713766c35c51fbc2e430ef6956f66c6a2614f43f85800c319e7f72096.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..60a97cf21fa2ce144781d259d5f197b7a4c09e0e
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/c4853ff713766c35c51fbc2e430ef6956f66c6a2614f43f85800c319e7f72096.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:759c934d37b0917cd2a95bd30e28adfb45cce45d1b850621e2d15c5f9b9d03f0
+size 1145
diff --git a/data/2025/2504_04xxx/2504.04158/images/c96f3d1612cc807bbae6b7d3a4d54788baface04ea76b2c13e72127c0a9b2776.jpg b/data/2025/2504_04xxx/2504.04158/images/c96f3d1612cc807bbae6b7d3a4d54788baface04ea76b2c13e72127c0a9b2776.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..88e8c74fdcb2b14578a5eb41047f31d50598f755
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/c96f3d1612cc807bbae6b7d3a4d54788baface04ea76b2c13e72127c0a9b2776.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eca174c7c7f610f190bdbddb837b19284829c3a93f054e53258be8aa176235eb
+size 28093
diff --git a/data/2025/2504_04xxx/2504.04158/images/cb056fa7fdbebf9868edbe0ad75e2c0e012de03e1875cafc854b2810b3dc02d3.jpg b/data/2025/2504_04xxx/2504.04158/images/cb056fa7fdbebf9868edbe0ad75e2c0e012de03e1875cafc854b2810b3dc02d3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..09e8067ee7b6baed8f0160fd8893d0e4e0d86097
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/cb056fa7fdbebf9868edbe0ad75e2c0e012de03e1875cafc854b2810b3dc02d3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3d918afc18f19e34d56cc90b6883f1ba38b3056f41483bcf7b3532c0def51cd
+size 256681
diff --git a/data/2025/2504_04xxx/2504.04158/images/cb91bb44da171cea19e4d652f14472183ab1a037dceea7015689eaf439a2247e.jpg b/data/2025/2504_04xxx/2504.04158/images/cb91bb44da171cea19e4d652f14472183ab1a037dceea7015689eaf439a2247e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..3ffd3b1d806aa42e4bcf4d2c8497fea3763e0972
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/cb91bb44da171cea19e4d652f14472183ab1a037dceea7015689eaf439a2247e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e906b8b94a4e77800fb59acf402c310f2edff31f16b0194168ce82c730ee997
+size 198807
diff --git a/data/2025/2504_04xxx/2504.04158/images/d14d0ce98a4d206614416c2089a3decdbad38fd58331486c17486cc99c111ac5.jpg b/data/2025/2504_04xxx/2504.04158/images/d14d0ce98a4d206614416c2089a3decdbad38fd58331486c17486cc99c111ac5.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..d02729805142eb15204bac002898cce153f94b23
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/d14d0ce98a4d206614416c2089a3decdbad38fd58331486c17486cc99c111ac5.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10b325a55aef11fb8d44b10327a59ca316e7854631afa250aea02797a8a5aee9
+size 5226
diff --git a/data/2025/2504_04xxx/2504.04158/images/d5950900eebc75f3811e0473fb96990fe04ce7e5528cde9a60accb30aaebde50.jpg b/data/2025/2504_04xxx/2504.04158/images/d5950900eebc75f3811e0473fb96990fe04ce7e5528cde9a60accb30aaebde50.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..f2a03d09429da41671681776c6992c81ce1deaab
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/d5950900eebc75f3811e0473fb96990fe04ce7e5528cde9a60accb30aaebde50.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aab933c72e2b98552a3bf0585369390e132fdaf69863649bdc1d325f7281c6e3
+size 95508
diff --git a/data/2025/2504_04xxx/2504.04158/images/e267ae823081c26f35980bf4508cf0c59d2aeb058e5d96797868cd0a40760317.jpg b/data/2025/2504_04xxx/2504.04158/images/e267ae823081c26f35980bf4508cf0c59d2aeb058e5d96797868cd0a40760317.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..b27dafc52b35f4b34c2ae25115cbbd0afe321620
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/e267ae823081c26f35980bf4508cf0c59d2aeb058e5d96797868cd0a40760317.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5188785a42d6ff41e95e531a7a861342a2ad4ed99475f4fe68164c06595b666c
+size 194211
diff --git a/data/2025/2504_04xxx/2504.04158/images/e3cf920182195ada7555a9b4f1b8abbf4999a4a80e5fb0df7cf67139c28b9385.jpg b/data/2025/2504_04xxx/2504.04158/images/e3cf920182195ada7555a9b4f1b8abbf4999a4a80e5fb0df7cf67139c28b9385.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ae3d6194d8d7199af86d1d947b35d3fa72a7ebfc
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/e3cf920182195ada7555a9b4f1b8abbf4999a4a80e5fb0df7cf67139c28b9385.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73d70b438fea267de8c605dcd9ed29dcacccc8478685a602a4c04ccc152ee0f9
+size 1540
diff --git a/data/2025/2504_04xxx/2504.04158/images/e81eb76098f90ddbcc64798aa45bc4836e2b2a1e67361ad4d5cc13d53950c11e.jpg b/data/2025/2504_04xxx/2504.04158/images/e81eb76098f90ddbcc64798aa45bc4836e2b2a1e67361ad4d5cc13d53950c11e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..ffbb57445cae230c997818ed0d052b09bcb2d07f
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/e81eb76098f90ddbcc64798aa45bc4836e2b2a1e67361ad4d5cc13d53950c11e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf694ced52e2691194c77ce985b0603e68f499b70739b488d87e72dccdba0994
+size 1457
diff --git a/data/2025/2504_04xxx/2504.04158/images/ed0a15f555772820149db0671eb9d687ea9446279705429b84b6d88753b2f9ff.jpg b/data/2025/2504_04xxx/2504.04158/images/ed0a15f555772820149db0671eb9d687ea9446279705429b84b6d88753b2f9ff.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..a19d7e60bdd66d48c4468f5b940c7eca6b31844f
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/ed0a15f555772820149db0671eb9d687ea9446279705429b84b6d88753b2f9ff.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf56641ba4d534c1483da0944316881e59870010085a537471867f79f1a1d0a3
+size 24777
diff --git a/data/2025/2504_04xxx/2504.04158/images/eda4a2ab495e8987b85b90c931c902aecf9588a2d577c2bb5b1bfb577136939f.jpg b/data/2025/2504_04xxx/2504.04158/images/eda4a2ab495e8987b85b90c931c902aecf9588a2d577c2bb5b1bfb577136939f.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..485f4b88cd482b8cdd0219e8f89beeccf675f8bb
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/eda4a2ab495e8987b85b90c931c902aecf9588a2d577c2bb5b1bfb577136939f.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e06401750c5361b503a64c4973d32491ff8bd585284f9ef2680d81ed3610574a
+size 5736
diff --git a/data/2025/2504_04xxx/2504.04158/images/ede7a4f309f3bcad5a30c932166e1189ac557ae03f9c4f5bf1590c35096c7eb0.jpg b/data/2025/2504_04xxx/2504.04158/images/ede7a4f309f3bcad5a30c932166e1189ac557ae03f9c4f5bf1590c35096c7eb0.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..cf72de90fe106944840223869ec4b8363484daca
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/ede7a4f309f3bcad5a30c932166e1189ac557ae03f9c4f5bf1590c35096c7eb0.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff1aee7616fff4d139095e59329d538ad8f090118f5f7bc9be12cb3dd8ab7403
+size 4029
diff --git a/data/2025/2504_04xxx/2504.04158/images/eedb1f16574b3946859a37dfc7f1e716129cc52a36878050a479074c6fdc306e.jpg b/data/2025/2504_04xxx/2504.04158/images/eedb1f16574b3946859a37dfc7f1e716129cc52a36878050a479074c6fdc306e.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..004b0bb49411bf3e8fc5182abaa79fa114d600ef
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/eedb1f16574b3946859a37dfc7f1e716129cc52a36878050a479074c6fdc306e.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86d2fc87cbc2784b6052f6953eada26a976a74bae7ab510e44297315017d6bd8
+size 335233
diff --git a/data/2025/2504_04xxx/2504.04158/images/f6c54bed650c992c605de2fa49637cefd6d3d1ff0e898590937ff0e9482905c1.jpg b/data/2025/2504_04xxx/2504.04158/images/f6c54bed650c992c605de2fa49637cefd6d3d1ff0e898590937ff0e9482905c1.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..dd523f29f8a2f5f087afbef6cec49f7485d4dc43
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/f6c54bed650c992c605de2fa49637cefd6d3d1ff0e898590937ff0e9482905c1.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:888730f92fdeb2e2b3bd01af1dba3942ec86233b794c7ab8815254790140aef6
+size 4640
diff --git a/data/2025/2504_04xxx/2504.04158/images/fb35655a752a3471f4a7966d044e73a0dc4c4d45f32e7ffd6704ea3beb55ddb3.jpg b/data/2025/2504_04xxx/2504.04158/images/fb35655a752a3471f4a7966d044e73a0dc4c4d45f32e7ffd6704ea3beb55ddb3.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..2563801ea644d3f7b0446f67e90ebd76bb23905e
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/fb35655a752a3471f4a7966d044e73a0dc4c4d45f32e7ffd6704ea3beb55ddb3.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fb12cfcaaccabda6461ab304c68506bef3415e9b9e0c8d87703bb203a8d10e0
+size 50184
diff --git a/data/2025/2504_04xxx/2504.04158/images/fd8c76b8464f35d32e0c3a8115eb5834370dd1176aec8322bf677aa7f13b38c8.jpg b/data/2025/2504_04xxx/2504.04158/images/fd8c76b8464f35d32e0c3a8115eb5834370dd1176aec8322bf677aa7f13b38c8.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..94ec4e663ac016559eee570618fcd4f0a68c8b47
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/images/fd8c76b8464f35d32e0c3a8115eb5834370dd1176aec8322bf677aa7f13b38c8.jpg
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10d9f83931f2dc5dff7f6ee3b4585253310dd66dae370ff31fb452abb37835a4
+size 474860
diff --git a/data/2025/2504_04xxx/2504.04158/layout.json b/data/2025/2504_04xxx/2504.04158/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..b84b3bb9651618fab9166b7a47810a693a80a54f
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04158/layout.json
@@ -0,0 +1,17823 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 160,
+ 110,
+ 483,
+ 145
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 160,
+ 110,
+ 483,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 110,
+ 483,
+ 145
+ ],
+ "type": "text",
+ "content": "JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "text",
+ "content": "Yunlong "
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Lin}^{1*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "text",
+ "content": " Zixu Lin"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "inline_equation",
+ "content": "^{1*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "text",
+ "content": " Haoyu Chen"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "inline_equation",
+ "content": "^{2*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "text",
+ "content": " Panwang Pan"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "inline_equation",
+ "content": "^{3*}"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "text",
+ "content": " Chenxin Li"
+ },
+ {
+ "bbox": [
+ 107,
+ 167,
+ 504,
+ 181
+ ],
+ "type": "inline_equation",
+ "content": "^{6}"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 158,
+ 182,
+ 462,
+ 196
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 158,
+ 182,
+ 462,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 158,
+ 182,
+ 462,
+ 196
+ ],
+ "type": "text",
+ "content": "Sixiang Chen² Yeying Jin⁴ Wenbo Li⁵† Xinghao Ding¹"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 118,
+ 196,
+ 492,
+ 210
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 196,
+ 492,
+ 210
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 196,
+ 492,
+ 210
+ ],
+ "type": "text",
+ "content": "| Task | Tools | Model Description |
| Super-resolution | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Denoising | SCUnet [94] | Hybrid UNet-based model combining convolution and transformer blocks, designed for robust denoising under diverse real-world noise conditions. |
| Compression artifact removal | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently. |
| Deblurring | StableSR-turbo [70] | Utilizes pre-trained diffusion models with a time-aware encoder for high-quality super-resolution, deblurring, and artifact removal. |
| Real-ESRGAN [72] | Fast GAN for super-resolution, deblurring, and artifact removal, handling complex real-world degradations efficiently.. |
| Deraining | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| UDR-S2Former [8] | An uncertainty-aware transformer model for rain streak removal. |
| Img2img-turbo-rain [52] | Efficient model based on SD-turbo, designed for fast and effective rain removal in real-world images. |
| Raindrop removal | IDT [82] | Transformer-based model for de-raining and raindrop removal. |
| Dehazing | RIDCP [81] | Efficient dehazing model utilizing high-quality codebook priors to handle complex real-world haze. |
| KANet [21] | Efficient dehazing network using a localization-and-removal pipeline to handle complex real-world hazy. |
| Desnowing | Img2img-turbo-snow [52] | Efficient model for removing snow artifacts while preserving natural scene details. |
| Snowformer [11] | Transformer-based model for removing snowflakes while preserving natural scene details. |
| Low-light enhancement | Img2img-turbo-night [52] | Fast and efficient model based on SD-turbo, designed for low-light enhancement in real-world scenarios. |
| HVI-CIDNet [83] | Lightweight transformer for low-light and exposure correction, enhancing both image quality and downstream vision tasks efficiently. |
| LightenDiff [27] | Diffusion-based framework for low-light enhancement, leveraging Retinex theory and latent-space decomposition for high-quality unsupervised restoration. |
",
+ "image_path": "eedb1f16574b3946859a37dfc7f1e716129cc52a36878050a479074c6fdc306e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 55,
+ 621,
+ 295,
+ 681
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 621,
+ 295,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 621,
+ 295,
+ 681
+ ],
+ "type": "text",
+ "content": "come from diverse sources, including the aforementioned autonomous driving datasets. Additionally, to enhance the generalizability of JarvisIR in natural contexts, we incorporated natural adverse weather scenes from internet and public datasets [5, 34, 39, 42, 47, 57, 87, 99]."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 314,
+ 620,
+ 476,
+ 634
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 620,
+ 476,
+ 634
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 620,
+ 476,
+ 634
+ ],
+ "type": "text",
+ "content": "9.2. Details of degradation library"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 313,
+ 642,
+ 555,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 642,
+ 555,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 642,
+ 555,
+ 715
+ ],
+ "type": "text",
+ "content": "As described in Sec 3.1 of the manuscript, we simulate realistic adverse weather scenarios—rainy, nighttime, snowy, and foggy conditions—by customizing a degradation library developed with physical models and image transformation techniques to synthesize degraded images. In this section, we detail our degradation implementations, cov"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 311,
+ 742
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 311,
+ 742
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 311,
+ 742
+ ],
+ "type": "text",
+ "content": "15"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 14
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 59,
+ 78,
+ 545,
+ 450
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 59,
+ 78,
+ 545,
+ 450
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 59,
+ 78,
+ 545,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 59,
+ 78,
+ 545,
+ 450
+ ],
+ "type": "image",
+ "image_path": "71a63f46812d2f4fe58ba8bc6a43938249dabe1c45d6bd32c8ceaa33a43536fc.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 55,
+ 459,
+ 555,
+ 495
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 459,
+ 555,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 459,
+ 555,
+ 495
+ ],
+ "type": "text",
+ "content": "Figure 8. Adverse weather scene simulator. To simulate realistic adverse weather scenarios, including rainy, nighttime, snowy, and foggy, we customized the degradation library developed using physical models and image transformation techniques to synthesize degraded images."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 55,
+ 505,
+ 298,
+ 554
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 505,
+ 298,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 505,
+ 298,
+ 554
+ ],
+ "type": "text",
+ "content": "ering the principles, formulas, and severity setups for the Night Scene Simulator, Fog Scene Simulator, Rain Scene Simulator, and Snow Scene Simulator. Examples for each implementation are provided in Figure 11."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "text",
+ "content": "Night Scene Simulator. Inspired by the work of [19], we employ a low-light degradation transform to synthesize realistic low-light images, denoted as "
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\text{night}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "text",
+ "content": ", as illustrated in Figure 8. Specifically, we first convert the daytime image "
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "I_{\\text{day}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "text",
+ "content": " into RAW data using the sRGB "
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\rightarrow"
+ },
+ {
+ "bbox": [
+ 55,
+ 570,
+ 298,
+ 715
+ ],
+ "type": "text",
+ "content": " RAW process [3]. Next, we linearly attenuate the RAW image and introduce Shot and Read (S&R) noise, which is commonly observed in camera imaging systems [3]. Finally, we apply the Image Signal Processing (ISP) pipeline to convert the low-light sensor data back into sRGB format. Additionally, we incorporate flare degradation using flare templates from the Flare7K++ [20] dataset. The complete low-light degrada"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 313,
+ 506,
+ 453,
+ 519
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 506,
+ 453,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 506,
+ 453,
+ 519
+ ],
+ "type": "text",
+ "content": "tion transform "
+ },
+ {
+ "bbox": [
+ 313,
+ 506,
+ 453,
+ 519
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\text{night}}"
+ },
+ {
+ "bbox": [
+ 313,
+ 506,
+ 453,
+ 519
+ ],
+ "type": "text",
+ "content": " is given by:"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 313,
+ 531,
+ 558,
+ 555
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 531,
+ 558,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 531,
+ 558,
+ 555
+ ],
+ "type": "interline_equation",
+ "content": "T _ {\\text {n i g h t}} \\left(\\mathrm {I} _ {\\text {d a y}}\\right) = T _ {I S P} \\left(T _ {s R G B \\rightarrow R A W} \\left(\\mathrm {I} _ {\\text {d a y}}\\right) + \\mathrm {I} _ {\\text {n o i s e}}\\right) + \\mathrm {I} _ {\\text {f l a r e}}, \\tag {8}",
+ "image_path": "a5c8902fafce1da343f0eebbc19d491c688868f6f7fd94729f9255756a759bd7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 313,
+ 555,
+ 555,
+ 616
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 555,
+ 555,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 555,
+ 555,
+ 616
+ ],
+ "type": "text",
+ "content": "which generates a degraded image "
+ },
+ {
+ "bbox": [
+ 313,
+ 555,
+ 555,
+ 616
+ ],
+ "type": "inline_equation",
+ "content": "I_{day}"
+ },
+ {
+ "bbox": [
+ 313,
+ 555,
+ 555,
+ 616
+ ],
+ "type": "text",
+ "content": " that closely resembles a dark nighttime scene. Furthermore, we use an online dynamic degradation process. It applies randomized parameter combinations, as defined in Equation 8, to simulate diverse nighttime driving conditions."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "text",
+ "content": "Fog Scene Simulator. Inspired by RIDCP [81], we design a foggy image degradation transform, denoted as "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\mathrm{fog}}"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "text",
+ "content": ", to synthesize realistic hazy images, as shown in Figure 8. Specifically, we simulate fog by introducing transmission maps "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "t(x)"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "text",
+ "content": " using depth estimation algorithms (e.g., Depth anything V2 [84]), combined with exponential attenuation "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "e^{\\beta d(x)}"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\beta"
+ },
+ {
+ "bbox": [
+ 313,
+ 617,
+ 556,
+ 715
+ ],
+ "type": "text",
+ "content": " controls haze density within the range [0.3, 1.5]. Additionally, poor lighting conditions are"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 312,
+ 742
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 312,
+ 742
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 312,
+ 742
+ ],
+ "type": "text",
+ "content": "16"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 15
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": "modeled by applying a brightness adjustment factor "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "inline_equation",
+ "content": "\\gamma \\in [1.5,3.0]"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": ", Gaussian noise "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{N}"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": ", and atmospheric light variation "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "inline_equation",
+ "content": "A + \\Delta A"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "inline_equation",
+ "content": "\\Delta A"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": " is sampled from "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "inline_equation",
+ "content": "[-0.025,0.025]"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": ". To further enhance realism, JPEG compression artifacts are introduced by applying JPEG "
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "inline_equation",
+ "content": "(\\cdot)"
+ },
+ {
+ "bbox": [
+ 55,
+ 72,
+ 294,
+ 145
+ ],
+ "type": "text",
+ "content": " to the degraded image. The complete foggy image synthesis process is defined as:"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 58,
+ 148,
+ 295,
+ 178
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 58,
+ 148,
+ 295,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 58,
+ 148,
+ 295,
+ 178
+ ],
+ "type": "interline_equation",
+ "content": "T _ {\\text {f o g}} \\left(I _ {\\text {d a y}}\\right) = \\operatorname {J P E G} \\left(\\mathcal {P} \\left(I _ {\\text {d a y}} ^ {\\gamma} + \\mathcal {N}, e ^ {\\beta d (x)}, A + \\Delta A\\right)\\right), \\tag {9}",
+ "image_path": "3ce7947a2594a77b7e6229dd1b8f04d4a9a1b99c63bb6ba190b22f88bd9ea21c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{P}"
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "content": " represents the hazy image formation process, "
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "inline_equation",
+ "content": "I_{\\mathrm{day}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "content": " is the clean image, and "
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "inline_equation",
+ "content": "d(x)"
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "content": " is the estimated depth map. The variable "
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "inline_equation",
+ "content": "x"
+ },
+ {
+ "bbox": [
+ 55,
+ 178,
+ 294,
+ 249
+ ],
+ "type": "text",
+ "content": " refers to the spatial coordinates of the image. This dynamic degradation process is designed to operate online with randomized parameters, simulating diverse real-world foggy conditions."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 55,
+ 250,
+ 294,
+ 321
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 250,
+ 294,
+ 321
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 250,
+ 294,
+ 321
+ ],
+ "type": "text",
+ "content": "Rain Scene Simulator. Inspired by PGDGN [54], we introduce a rain degradation transform, denoted as "
+ },
+ {
+ "bbox": [
+ 55,
+ 250,
+ 294,
+ 321
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\mathrm{rain}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 250,
+ 294,
+ 321
+ ],
+ "type": "text",
+ "content": ", to generate realistic rainy images (Figure 8). This transform synthesizes rainy images by combining a disentangled clean image with a physics-based rain rendering model. The degradation process is formulated as:"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 97,
+ 327,
+ 294,
+ 340
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 97,
+ 327,
+ 294,
+ 340
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 97,
+ 327,
+ 294,
+ 340
+ ],
+ "type": "interline_equation",
+ "content": "T _ {\\text {r a i n}} \\left(I _ {\\text {d a y}}\\right) = W _ {\\text {M o d}} \\left(G \\left(I _ {\\text {d a y}}\\right), \\hat {w}, z\\right), \\tag {10}",
+ "image_path": "917405f323bfed15e07d97060629637b7c1a50fd3d1629b96986c5c02ab0c32b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "I_{\\mathrm{day}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " is the clean image, "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "G(I_{\\mathrm{day}})"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " represents the disentangled base image, and "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "W_{\\mathrm{Mod}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " is the rain rendering model. "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "W_{\\mathrm{Mod}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " incorporates parameters "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{w} = \\{\\hat{w}_d, \\hat{w}_{nd}\\}"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": ", with "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{w}_d"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " controlling differentiable aspects such as raindrop size and streak density, and "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{w}_{nd}"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " addressing nondifferentiable properties. The term "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "z"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " introduces stochastic noise for variability in rain effects. This process applies "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "W_{\\mathrm{Mod}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " to add realistic raindrop occlusions, rain streaks, and scene wetness to the disentangled image. "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "G(I_{\\mathrm{day}})"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": ", generating a visually plausible rainy image "
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\mathrm{rain}}(I_{\\mathrm{day}})"
+ },
+ {
+ "bbox": [
+ 55,
+ 344,
+ 295,
+ 474
+ ],
+ "type": "text",
+ "content": " with controlled and diverse effects."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 55,
+ 475,
+ 295,
+ 559
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 475,
+ 295,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 475,
+ 295,
+ 559
+ ],
+ "type": "text",
+ "content": "Snow Scene Simulator. Building on the img2img-turbo model [52], we introduce a snow transformation, denoted as "
+ },
+ {
+ "bbox": [
+ 55,
+ 475,
+ 295,
+ 559
+ ],
+ "type": "inline_equation",
+ "content": "T_{\\mathrm{snow}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 475,
+ 295,
+ 559
+ ],
+ "type": "text",
+ "content": ", to generate realistic snowy images from daytime inputs. This process uses the SD-Turbo model with textual conditioning. It synthesizes snowy scenes by combining the input image with a latent diffusion-based generator and a textual prompt. The snow transformation is formulated as:"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 78,
+ 564,
+ 294,
+ 578
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 564,
+ 294,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 564,
+ 294,
+ 578
+ ],
+ "type": "interline_equation",
+ "content": "T _ {\\text {s n o w}} \\left(I _ {\\text {d a y}}, C _ {\\text {s n o w}}\\right) = G _ {\\text {s n o w}} \\left(I _ {\\text {d a y}}, C _ {\\text {s n o w}}\\right), \\tag {11}",
+ "image_path": "2782c78e07bdd618db0f57c0df1cbc2cedfb60ec2c542a163c746b4b6c60921e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "I_{\\mathrm{day}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": " is the daytime input image, "
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "C_{\\mathrm{snow}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": " is the textual condition (e.g., \"driving in the heavy snow\"), and "
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "G_{\\mathrm{snow}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": " represents the generator. By employing LoRA adapters and skip connections, the generator enables precise control over scene characteristics while maintaining the structural integrity of the input image. This process applies "
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "G_{\\mathrm{snow}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": " to infuse the daytime image "
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "I_{\\mathrm{day}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": " with snowy features, guided by the contextual information in "
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "C_{\\mathrm{snow}}"
+ },
+ {
+ "bbox": [
+ 55,
+ 582,
+ 295,
+ 713
+ ],
+ "type": "text",
+ "content": ". The resulting synthetic image aligns closely with the visual expectations of a snowy environment while maintaining consistency with the original scene's structure."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 315,
+ 71,
+ 409,
+ 83
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 71,
+ 409,
+ 83
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 71,
+ 409,
+ 83
+ ],
+ "type": "text",
+ "content": "10. More ablation"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 313,
+ 91,
+ 553,
+ 163
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 91,
+ 553,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 91,
+ 553,
+ 163
+ ],
+ "type": "text",
+ "content": "To thoroughly investigate the proposed JarvisIR, we conducted an extensive array of ablation studies on the CleanBench-Real dataset. Four non-reference metrics are used for assessment: MUSIQ [35], MANIQA [86], CLIP-IQA+ [69], LIQE [95]. The specific elements of these studies are further expounded in the sections that follow."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 314,
+ 172,
+ 468,
+ 184
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 172,
+ 468,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 172,
+ 468,
+ 184
+ ],
+ "type": "text",
+ "content": "10.1. MRRHF vs. vanilla RRHF"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 313,
+ 190,
+ 555,
+ 370
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 190,
+ 555,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 190,
+ 555,
+ 370
+ ],
+ "type": "text",
+ "content": "We evaluate the effectiveness of our proposed MRRHF by comparing it with vanilla RRHF [93]. The reward and diversity metrics over training iterations are illustrated in Table 5. Fine-tuning JarvisIR with MRRHF significantly improves the average values of both reward and diversity by 0.19 and 3.43, respectively, compared to using RRHF. The degradation in diversity and reward when using vanilla RRHF results from its offline sample generation strategy. As discussed in Sec. 5 in the manuscript, this strategy confines its generated samples to the finite sample space created by the SFT model using diverse beam search [68]. In contrast, our MRRHF employs a hybrid sample generation strategy and entropy regularization, providing sufficient sample exploration space to achieve globally optimal results."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 315,
+ 380,
+ 553,
+ 403
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 315,
+ 380,
+ 553,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 315,
+ 380,
+ 553,
+ 403
+ ],
+ "type": "text",
+ "content": "10.2. Sample generation strategy and entropy regularization"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 313,
+ 409,
+ 554,
+ 554
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 409,
+ 554,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 409,
+ 554,
+ 554
+ ],
+ "type": "text",
+ "content": "In our manuscript, we examine the effects of the sample generation strategy and entropy regularization on the MR-RHF tuning process, focusing on reward scores and response diversity. This section provides further evidence of the effectiveness of our hybrid sample generation strategy and entropy regularization. Specifically, as shown in Table 5, we assess their impact on performance using the CleanBench-Real validation set. The results demonstrate that our hybrid sampling approach and entropy regularization not only enhance training stability and facilitate high-quality exploration of the optimization space but also significantly improve testing performance."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 314,
+ 562,
+ 553,
+ 587
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 562,
+ 553,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 562,
+ 553,
+ 587
+ ],
+ "type": "text",
+ "content": "10.3. Effectiveness of differentiated contrast weights"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "text",
+ "content": "In Equation 4 of our manuscript, we refine the original ranking loss [93] by introducing differentiated contrast weights, expressed as "
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "L_{\\mathrm{rank}} = \\sum_{s_i < s_j} (s_j - s_i) \\max(0, p_i - p_j)"
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "text",
+ "content": ". The term "
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "(s_j - s_i)"
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "text",
+ "content": " represents the differentiated contrast weights. We compare this with the original ranking loss "
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{L}_{\\mathrm{rank}} = \\sum_{s_i < s_j} \\max(0, p_i - p_j)"
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "text",
+ "content": ". Table 5 presents the reward and diversity metrics over training iterations. When the ranking loss is applied without differentiated contrast weights "
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{L}_{\\mathrm{rank}}"
+ },
+ {
+ "bbox": [
+ 313,
+ 592,
+ 555,
+ 714
+ ],
+ "type": "text",
+ "content": " the average values of both reward and diversity decrease by 0.14 and 3.93, respectively, compared"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 311,
+ 742
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 311,
+ 742
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 732,
+ 311,
+ 742
+ ],
+ "type": "text",
+ "content": "17"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 16
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 69,
+ 125,
+ 547,
+ 269
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 55,
+ 67,
+ 555,
+ 124
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 67,
+ 555,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 67,
+ 555,
+ 124
+ ],
+ "type": "text",
+ "content": "Table 5. Ablation studies on tuning paradigm, differentiated contrast weights, different sample generation strategies, and entropy regularization. The \"Reward\" represents the average reward scores obtained during alignment tuning, spanning from -1 to 1. A negative score indicates a penalty, while a positive score represents a reward. The \"Diversity\" reflects the average number of unique responses produced during the training process. Additionally, we evaluate performance on the CleanBench-Real validation set using four non-reference metrics: MUSIQ, MANIQA, CLIP-IQA+, and LIQE. The reported values represent the average performance across all tested scenes."
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 69,
+ 125,
+ 547,
+ 269
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 125,
+ 547,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 125,
+ 547,
+ 269
+ ],
+ "type": "table",
+ "html": "| # | Instruction |
| 1 | Please evaluate the image's quality comprehensively and provide your insights. Additionally, outline a step-by-step restoration strategy, specifying the sequence of tasks and model choices for each step. The available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement. The available restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 2 | Analyze the quality of the image comprehensively and provide your insights. Furthermore, propose a restoration strategy by detailing each task and model choice sequentially. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement. The restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 3 | Assess the overall quality of the image and provide a detailed evaluation. Then, design a step-by-step restoration process, specifying tasks and model choices. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 4 | Perform a comprehensive evaluation of the image quality and explain your observations. Additionally, develop a step-by-step restoration plan, identifying tasks and model choices. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 5 | Conduct a thorough analysis of the image's quality and provide your insights. Subsequently, create a restoration strategy step by step, specifying the tasks and model choices. The tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 6 | Evaluate the quality of the image comprehensively and outline your findings. Moreover, formulate a sequential restoration plan, detailing tasks and model selections. Available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 7 | Provide a detailed assessment of the image's quality and share your observations. Then, create a restoration strategy in a step-by-step manner, specifying tasks and models. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and models include StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 8 | Examine the image comprehensively and describe its quality in detail. Additionally, outline a step-wise restoration process, specifying the tasks and models for each stage. Tasks available are super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 9 | Review the image quality comprehensively and present your findings. Furthermore, propose a restoration plan step by step, detailing the tasks and models used. Available tasks are super-resolution, denoising, dehazing, ..., low-light enhancement, and restoration models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
| 10 | Perform a detailed evaluation of the image and provide your insights. Then, construct a step-by-step restoration strategy, specifying the sequence of tasks and the corresponding models. The available tasks include super-resolution, denoising, dehazing, ..., low-light enhancement, and models are StableSR-turbo: (model description), SCUnet: (model description), ..., RIDCP: (model description). |
",
+ "image_path": "fd8c76b8464f35d32e0c3a8115eb5834370dd1176aec8322bf677aa7f13b38c8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 311,
+ 742
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 311,
+ 742
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 311,
+ 742
+ ],
+ "type": "text",
+ "content": "24"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 23
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 78,
+ 118,
+ 531,
+ 685
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 165,
+ 97,
+ 445,
+ 109
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 165,
+ 97,
+ 445,
+ 109
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 165,
+ 97,
+ 445,
+ 109
+ ],
+ "type": "text",
+ "content": "Table 9. Responses generated by GPT-4V using the self-instruct strategy [73]"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 78,
+ 118,
+ 531,
+ 685
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 118,
+ 531,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 118,
+ 531,
+ 685
+ ],
+ "type": "table",
+ "html": "| # | Response |
| 1 | The image depicts an outdoor scene viewed through a wet windshield, with water droplets significantly obscuring the view. The evaluated image exhibits severe distortion caused by the water droplets, resulting in a loss of sharpness and clarity. Details of the background, such as the car ahead and surrounding structures, are blurred and difficult to discern. The image also suffers from uneven focus and lacks proper visibility. |
| 2 | This image portrays an outdoor scene as seen through a wet windshield, where water droplets heavily obstruct the view. The distortion caused by these droplets leads to a noticeable reduction in sharpness and clarity. Background elements, like the car in front and nearby structures, appear heavily blurred and challenging to distinguish. The image further suffers from uneven focus, reducing overall visibility. |
| 3 | An outdoor setting is captured in this image, viewed through a wet windshield. The water droplets significantly obscure the scene, causing severe distortion. Sharpness and clarity are greatly reduced, and background features, such as the car ahead and surrounding buildings, are blurred and hard to recognize. Uneven focus further detracts from the image's overall visibility. |
| 4 | This image showcases an outdoor scene as observed through a wet windshield, with water droplets causing significant obstruction. The distortion leads to substantial clarity and sharpness loss. Details in the background, including the car ahead and nearby structures, are blurred and hard to identify. Additionally, the image has uneven focus, which impairs its overall visibility. |
| 5 | The image illustrates an outdoor view observed through a wet windshield, where water droplets obscure much of the scene. The distortion caused by the droplets results in a notable reduction in sharpness and clarity. Background details, such as the car ahead and surrounding structures, appear blurry and indistinct. Uneven focus further degrades the visibility of the image. |
| 6 | In this image, an outdoor setting is depicted as seen through a wet windshield, with water droplets heavily obscuring the view. The distortion from the droplets results in a considerable loss of sharpness and clarity. Background elements, like the car ahead and nearby buildings, are blurred and hard to discern. The uneven focus of the image adds to its overall lack of visibility. |
| 7 | This image shows an outdoor scene viewed through a wet windshield, with water droplets creating significant visual obstruction. The distortion caused by the droplets reduces the sharpness and clarity of the image. Background features, including the car ahead and the surrounding structures, are blurred and difficult to distinguish. The image also suffers from uneven focus, limiting its visibility. |
| 8 | An outdoor view is displayed in this image, captured through a wet windshield with water droplets blocking much of the scene. The distortion caused by the droplets results in reduced clarity and sharpness. Background details, such as the car in front and nearby structures, are blurred and indistinct. Uneven focus further compromises the visibility of the image. |
| 9 | The image depicts an outdoor scene seen through a wet windshield, where water droplets obscure much of the view. This distortion leads to a significant loss of sharpness and clarity. Background elements like the car ahead and nearby structures are blurred and challenging to distinguish. Uneven focus further contributes to the poor visibility of the image. |
| 10 | This image represents an outdoor scene viewed through a wet windshield, with water droplets obscuring the visual field. The distortion caused by the droplets significantly affects sharpness and clarity, making background features like the car ahead and surrounding structures appear blurred and indistinct. The uneven focus further reduces the overall visibility of the image. |
",
+ "image_path": "46e0fd7cfb79df6a4f73542600a61049ba8364fa75e365b59aaf3d2e9e3a372c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 310,
+ 742
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 310,
+ 742
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 732,
+ 310,
+ 742
+ ],
+ "type": "text",
+ "content": "25"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 24
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/data/2025/2504_04xxx/2504.04259/740753b2-ad81-48a6-b430-9d2ffd0735ec_content_list.json b/data/2025/2504_04xxx/2504.04259/740753b2-ad81-48a6-b430-9d2ffd0735ec_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..9d313c47bcf751b72463ea3abcfec71b154f7871
--- /dev/null
+++ b/data/2025/2504_04xxx/2504.04259/740753b2-ad81-48a6-b430-9d2ffd0735ec_content_list.json
@@ -0,0 +1,1394 @@
+[
+ {
+ "type": "text",
+ "text": "ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning",
+ "text_level": 1,
+ "bbox": [
+ 104,
+ 87,
+ 893,
+ 138
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Clemens C. Christoph $^{1\\dagger}$ , Maximilian Eberlein $^{1\\dagger}$ , Filippos Katsimalis $^{1\\dagger}$ , Arturo Roberti $^{1\\dagger}$ , Aristotelis Sympetheros $^{1\\dagger}$ , Michel R. Vogt $^{1\\dagger}$ , Davide Liconti $^{1}$ , Chenyu Yang $^{1}$ , Barnabas Gavin Cangan $^{1}$ , Ronan J. Hinchet $^{1}$ , Robert K. Katzschmann $^{1*}$",
+ "bbox": [
+ 112,
+ 157,
+ 859,
+ 209
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract—General-purpose robots should possess human-like dexterity and agility to perform tasks with the same versatility as us. A human-like form factor further enables the use of vast datasets of human-hand interactions. However, the primary bottleneck in dexterous manipulation lies not only in software but arguably even more in hardware. Robotic hands that approach human capabilities are often prohibitively expensive, bulky, or require enterprise-level maintenance, limiting their accessibility for broader research and practical applications. What if the research community could get started with reliable dexterous hands within a day? We present the open-source ORCA hand, a reliable and anthropomorphic 17-DoF tendon-driven robotic hand with integrated tactile sensors, fully assembled in less than eight hours and built for a material cost below 2,000 CHF. We showcase ORCA's key design features such as popping joints, auto-calibration, and tensioning systems that significantly reduce complexity while increasing reliability, accuracy, and robustness. We benchmark the ORCA hand across a variety of tasks, ranging from teleoperation and imitation learning to zero-shot sim-to-real reinforcement learning. Furthermore, we demonstrate its durability, withstanding more than 10,000 continuous operation cycles—equivalent to approximately 20 hours—without hardware failure, the only constraint being the duration of the experiment itself. Video is here: youtu.be/kUbPSYMmOds. Design files, source code, and documentation are available at sr1.ethz.ch/orcahand.",
+ "bbox": [
+ 81,
+ 233,
+ 491,
+ 561
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "I. INTRODUCTION",
+ "text_level": 1,
+ "bbox": [
+ 218,
+ 571,
+ 354,
+ 585
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Reproducing the intricate dexterity of the human hand has long been a central challenge in robotics [1], [2]. Although robotic grippers excel in industrial automation, their limited versatility makes them unsuitable for interaction with tools and objects designed for human hands [3], [4]. Consequently, extensive research has been devoted to developing anthropomorphic robotic hands and training them to solve complex manipulation tasks [5], [6]. However, compared to grippers, anthropomorphic hands require significantly more actuators, increasing the complexity of their assembly and control. In addition, good anthropomorphic hand hardware must be durable, repeatable, and versatile to be used in machine learning applications. Whether it is the sim2real gap in reinforcement learning (RL) or the accuracy of teleoperation in imitation learning (IL), bottlenecks in dexterous manipulation stem not only from software, but also, and maybe more importantly, from hardware limitations [7]–[9].",
+ "bbox": [
+ 81,
+ 590,
+ 490,
+ 847
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Tendon-driven robotic hands have been a focal point of robotics research since its early days [11]-[13]. Among",
+ "bbox": [
+ 83,
+ 848,
+ 488,
+ 878
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "$^{1}$ Soft Robotics Lab, IRIS, D-MAVT, ETH Zurich, Switzerland",
+ "$\\dagger$ Equal contribution.",
+ "*Corresponding author: rkk@ethz.ch"
+ ],
+ "bbox": [
+ 96,
+ 888,
+ 439,
+ 925
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/4b6b8ae9e031516a88624eb4e0c405e7d29a02b46042fbaa38296912da9df06e.jpg",
+ "image_caption": [
+ "Fig. 1: (A) The ORCA hand closely mimics its human counterpart with the same form factor, a bony structure, and silicone-cast skin. The ORCA hand is 3D-printed but incorporates joints designed to pop before breaking, making it resistant to overload-induced failures while retaining the advantages of bearing pinhole joints, such as stability and simple kinematics. (A1) Just before the joint pops. (A2) Applying pressure pops the joint into place and keeps it secure. (A3) Depicts our spool system, which enables manual retention without unscrewing the spools or tendons. (B) We show that our hand can be deployed in real-world settings by running our self-resetting imitation learning policy for over 7 hours before we decided to end the experiment. (C) Our reliability test reveals our hand's robustness and the high repeatability of joint movements."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 509,
+ 227,
+ 913,
+ 652
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "the various designs, the proprietary Shadow Hand [14] demonstrated impressive capabilities in dexterous manipulation tasks [15]. However, these hands cost over 100,000 CHF, require substantial maintenance [16] and are difficult to repair due to their proprietary and highly integrated designs. Other recent tendon-driven hand designs, such as the InMoov hand [17] and the DexHand [18], offer the advantages of being open source and low-cost. However, the",
+ "bbox": [
+ 504,
+ 805,
+ 913,
+ 928
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "aside_text",
+ "text": "arXiv:2504.04259v2 [cs.RO] 17 Sep 2025",
+ "bbox": [
+ 22,
+ 276,
+ 60,
+ 720
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/5134e3c491e37dccf44e14f3fdbdfabda7fc07fb2fc2b4d3ee3231a4e50c8004.jpg",
+ "image_caption": [
+ "Fig. 2: Versatility of the ORCA hand: (A)-(D) Teleoperation with ROKOKO [10] gloves. (A) Holding a pen (B) Using a drill, showing high dexterity. (C) Liquid pouring. (D) Grasping a cube: This picture illustrates how closely the ORCA hand resembles a human hand. (E) IL with walls and a slider for self-resetting. (F) Policies in simulation, such as rolling a ball, can be deployed zero-shot to the real world due to the ORCA hand's low joint errors."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 84,
+ 66,
+ 233,
+ 250
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/38cb7e77c0c79367fec280ed24dd497aa5606fc23634c0a827cde3e2e170570b.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 233,
+ 68,
+ 343,
+ 250
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/364a3dfedca85940a2e9dcb442264acd24618fd92a493bb23dd7cde11a16f043.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 344,
+ 68,
+ 496,
+ 250
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/a02ee3ec40125321ff1540bf9c516279ed35bdcbb35e46c0bf90177f6c635fb6.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 496,
+ 68,
+ 676,
+ 250
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/6e6215471510c69b27506674c3cf7d2a87ce7f61e9b5f4f87d0625de53996693.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 676,
+ 68,
+ 911,
+ 250
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "InMoov hand is limited in dexterity, while the DexHand is challenging to assemble, and neither has demonstrated real-world applicability in autonomous manipulation tasks.",
+ "bbox": [
+ 81,
+ 300,
+ 488,
+ 345
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Alternatives to tendon-driven hands are direct-driven hands, such as the Allegro Hand [19], priced at around 15,000 CHF, and the open-source LEAP hand [20], which is even more cost-effective under 2,000 CHF and requires only three hours of assembly for unprecedented levels of reliability. However, all direct-driven hand designs share limitations such as bulkiness, restricted form factors, and an inability to match the softness, form factor, and agility of a human hand. Embedding motors within the fingers increases inertia and limits power output, restricting the quick, dynamic, and forceful movements essential for human-like motion.",
+ "bbox": [
+ 81,
+ 345,
+ 488,
+ 511
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In this paper, we present the ORCA hand: a tendon-driven, dexterous, and anthropomorphic robotic hand with fully integrated tactile sensors. The ORCA hand is designed for reliability, simplicity, and versatility in a wide range of tasks. Key contributions of our integrated system include:",
+ "bbox": [
+ 81,
+ 512,
+ 488,
+ 588
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- An open source, 3D-printable design with a cost of less than 2,000 CHF, which can be assembled by a single person without prior experience in less than eight hours.",
+ "- A joint design that pops before it breaks, enhancing the durability of 3D-printed components and streamlining the assembly process.",
+ "- Auto-calibration enabled by tendon routing through the center of rotation. This minimizes joint position errors and increases repeatability.",
+ "- Fully integrated tactile sensors and sensor wiring, which can be produced in-house, offering a compact and modular solution."
+ ],
+ "bbox": [
+ 99,
+ 590,
+ 488,
+ 771
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "We demonstrate the hand's dexterity by teleoperating it to perform complex tasks that traditional robotic grippers cannot accomplish. Through a variety of reliability tests we demonstrate the ORCA hand's capability to offer exceptional reliability, durability, and consistent performance during tens of hours of operation. To showcase the ORCA hand's dexterity and accuracy we implement fine motor control tasks like in-hand object orientation. Leveraging the anthropomorphic design, we are also able to implement imitation learning tasks like the picking and placement of cubes and execute these",
+ "bbox": [
+ 81,
+ 775,
+ 490,
+ 926
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "autonomous tasks continuously for multiple hours without any human intervention on the hand hardware.",
+ "bbox": [
+ 504,
+ 301,
+ 911,
+ 330
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "II. SYSTEM DESIGN",
+ "text_level": 1,
+ "bbox": [
+ 635,
+ 342,
+ 784,
+ 356
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The ORCA hand design follows the requirements set above, namely dexterity and reliability at minimal complexity and cost. It comprises five fingers, including an opposable thumb and an actuated wrist (Figure 1-A), and is of a size similar to the average human hand [21]. The total weight of the hand is approximately $1.2\\mathrm{kg}$ . The fingers are mounted on a base that resembles the human palm containing the carpal and metacarpal bones. The palm is connected to the wrist mechanism that is mounted on the tower. The tower contains the motors and all auxiliary electronics, and is enclosed in a protective casing. The rest of this section describes the most important design features of the ORCA hand in detail.",
+ "bbox": [
+ 504,
+ 362,
+ 913,
+ 544
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "A. Tendon Actuation for Agility",
+ "text_level": 1,
+ "bbox": [
+ 504,
+ 556,
+ 725,
+ 570
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Each joint, except for the wrist joint, is actuated using two fishing lines (Nylon fibers braided into a $0.4\\mathrm{mm}$ diameter rope) under tension, here referred to as tendons. One tendon is responsible for flexion (flexor) and the other for extension (extensor). This decision was made based on the observation that a smaller form factor and lower finger inertia more closely mimic the nimbleness and dexterity of human hands compared to direct-driven hands. In addition, tendon actuation makes the ORCA hand independent of the choice of actuator, enabling actuation technologies other than electric motors to be used in the future, such as contracting artificial muscles.",
+ "bbox": [
+ 504,
+ 575,
+ 913,
+ 755
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Although tendon actuation offers many advantages, it also presents challenges such as friction build-up, wear, and slack over time, which can affect movement precision and longevity. We mitigate those challenges as follows:",
+ "bbox": [
+ 504,
+ 757,
+ 911,
+ 818
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- We avoid direct contact of the tendons with polylactic acid (PLA) by deflecting the tendons around smooth metal pins and rods, as depicted in Figure 3-C.",
+ "- We use Teflon tubes for nonlinear routing, e.g., from the bottom of the thumb to the wrist.",
+ "- Finally, tendons can be manually re-tensioned using a ratchet spool mechanism mounted on the motors,"
+ ],
+ "bbox": [
+ 522,
+ 820,
+ 911,
+ 926
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/6220ded5110b61f95110e4b02ca4dd030e39f0bc4994a1fd8bc10e29a1987705.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 86,
+ 66,
+ 486,
+ 294
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "image",
+ "img_path": "images/e181f221dc68e3d036837b2a64b71ae487cb2528fa325945bc62d8896ed11dbb.jpg",
+ "image_caption": [
+ "Fig. 3: (A) Naming convention of the joints. The thumb includes an additional degree of freedom. (B) Auto-calibration: The three-step process moves all joints to their respective limits and determines a mapping between motor and joint angles without any external sensors. (C) Routing of the DIP joint. The tendons are guided around metal pins, to reduce friction and eliminate wear over time. Moreover, they are always guided through the center of rotation for straightforward control at minimal slack."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 86,
+ 296,
+ 478,
+ 401
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "as shown in Figure 1-A3. The ratchet is attached to the top spool, allowing rotation in one direction while locking movement in the other. This design makes the ORCA hand user-friendly, as re-tensioning can be done in seconds without the need to unscrew the spool or tendon. The spool system quickly removes any slack that accumulates. In addition, tendons can be easily loosened, allowing joints to pop out for quick replacement of broken parts.",
+ "bbox": [
+ 114,
+ 500,
+ 488,
+ 635
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "B. Poppable Pin Joints",
+ "text_level": 1,
+ "bbox": [
+ 83,
+ 659,
+ 243,
+ 674
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Rolling contact joints [18], [22] have become a popular alternative to pinhole joints for 3D-printed hands due to their ability to dislocate instead of break. However, these mechanisms require ligaments that can potentially loosen over time and increase complexity. We introduce a pin joint design that allows the joints to \"pop\" out of place and dislocate instead of breaking when excessive radial and axial loads are applied (Figure 1-A1).",
+ "bbox": [
+ 81,
+ 683,
+ 488,
+ 804
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "This feature is achieved by placing the bearings in circular arc-shaped grooves, which hold them tightly under normal operating conditions, but allow them to dislocate in the event of a forceful collision. This mechanism combines the advantages of pinhole joints, such as axial stability and straightforward kinematics, with the robustness of ligament-based rolling contact joints, while being extremely quick and easy to assemble.",
+ "bbox": [
+ 81,
+ 805,
+ 490,
+ 926
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "C. Finger and Palm Design",
+ "text_level": 1,
+ "bbox": [
+ 506,
+ 66,
+ 700,
+ 80
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Fingers two to five (index to pinky) have three actuated joints that mimic those of the human finger: the Proximal Interphalangeal (PIP), the Metacarpophalangeal (MCP), and the Abduction (ABD) joint (Figure 3-A). The ORCA PIP joint corresponds to the human PIP joint, while the MCP and ABD joints together correspond to the human MCP joint which can perform both flexion/extension and abduction/adduction, which is necessary in various dexterous manipulation tasks [23]. The range of motion (RoM) for each joint is shown in Table I, based on human anatomy [21].",
+ "bbox": [
+ 504,
+ 84,
+ 913,
+ 237
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "table",
+ "img_path": "images/7afd4f70482b1c101b03f7b6a712ac0cf436088884cd1bcd4a71936d9c7371be.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "