igtsd/acad_dataset

ACAD: Amodal Concrete Aggregate Dataset

I Nyoman Gede Arya Astawa¹, I Gede Teguh Satya Dharma¹, I Nyoman Suardika², and Made Pradnyana Ambara¹

¹ Department of Information Technology, Politeknik Negeri Bali
² Department of Civil Engineering, Politeknik Negeri Bali
_{Correspondence to: arya_kmg@pnb.ac.id or teguh@pnb.ac.id}

Abstract

ACAD (Amodal Concrete Aggregate Dataset) is a procedurally generated synthetic image dataset with a resolution of 512 × 512 pixels, designed to support research on amodal instance segmentation for concrete aggregates. The dataset comprises 10,000 images organized into two controlled occlusion levels: low occlusion (3,000 images with a 10% overlap ratio) and high occlusion (7,000 images with overlap ratios of up to 55%). ACAD defines two object categories: concrete aggregates and reference coins for physical scale calibration. Each object instance is annotated with both a modal polygon, representing the visible region, and an amodal polygon, representing the complete object silhouette. The dataset generation pipeline is fully deterministic through seed-controlled procedural generation, incorporating graded particle size distributions, random object rotations, photometric augmentation, and soft directional shadows. Released under the CC BY 4.0 license, ACAD provides a large-scale, controlled, and readily usable training dataset without requiring costly manual annotation for concrete material analysis.

Dataset Structure

The dataset is organized into separate directories according to occlusion difficulty levels, facilitating efficient data loading.

Directory Layout

ACAD_Dataset/
├── low_occlusion/               # Low Occlusion Subset
│   ├── images/                  # Output Images JPG (prefix: lo_*)
│   ├── train.json               # COCO Annotation + Amodal (2.400 images)
│   ├── val.json                 # COCO Annotation + Amodal (300 images)
│   └── test.json                # COCO annotation + Amodal (300 images)
└── high_occlusion/              # High Occlusion Subset
    ├── images/                  # Output Images JPG (prefix: hi_*)
    ├── train.json               # COCO Annotation + Amodal (5.600 citra)
    ├── val.json                 # COCO Annotation + Amodal (700 citra)
    └── test.json                # COCO Annotation + Amodal (700 citra)

Quantitative Specification

Specification	low_occlusion	high_occlusion
Number of Images	3000 Images	7000 Images
Number of Aggregates	10 - 20 Objects	21 - 35 Objects
Number of Coins	1 - 2 Objects	1 - 2 Objects
Occlusion Ratio	~10%	~55%
Image Resolution & Format	512 × 512 Pixels (JPG)	512 × 512 Pixels (JPG)
Files	lo_000000.jpg to lo_002999.jpg	hi_000000.jpg to hi_006999.jpg
Data Split (80/10/10%)	Train (2400) / Val (300) / Test (300)	Train (5600) / Val (700) / Test (700)

Annotation Schema (COCO JSON + Amodal Extension)

Annotations are stored in the standard COCO JSON format, extended with a custom field for amodal annotations. The main components of the JSON file are as follows:

Categories

The dataset defines two object categories:

• aggregate (ID: 1): Concrete aggregate particles with varying shapes, sizes, and orientations.
• reference_coin (ID: 2): Reference coins used as physical scale references for real-world size measurements.

"categories": [
  { "id": 1, "name": "aggregate", "supercategory": "stone" },
  { "id": 2, "name": "reference_coin", "supercategory": "coin" }
]

Annotations

Each object instance in the dataset is annotated with the following attributes:

• bbox: Bounding box in the format [x, y, width, height] enclosing the object.
• segmentation: Polygon coordinates representing the visible region of the object (modal mask).
• amodal_segmentation: Polygon coordinates representing the complete object silhouette, including occluded regions (amodal mask).
• area: Object area in pixels, computed consistently from the complete silhouette defined by amodal_segmentation.
• category_id: Object category identifier (1 for aggregate, 2 for reference_coin).
• iscrowd: Always set to 0, indicating that all object instances are annotated individually.

Example Annotation Snippet

        {
            "id": 44,
            "image_id": 2,
            "category_id": 1,
            "bbox": [305, 248, 50, 67],
            "area": 2406,
            "iscrowd": 0,
            "segmentation": [
                [
                    325, 248, 325, 250, 323, 252, 323, 253, 321, 255, 321, 256,
                    320, 257, 320, 258, 319, 259, 319, 260, 318, 261, 318, 265,
                    319, 266, 319, 267, 320, 268, 320, 272, 321, 273, 321, 284,
                    322, 285, 322, 287, 321, 288, 321, 289, 322, 290, 322, 298,
                    321, 299, 321, 306, 320, 307, 321, 308, 321, 310, 323, 310,
                    324, 311, 329, 311, 330, 312, 331, 312, 332, 313, 334, 313,
                    335, 314, 336, 314, 337, 313, 340, 313, 342, 311, 343, 311,
                    344, 310, 345, 310, 347, 308, 348, 308, 350, 306, 351, 306,
                    353, 304, 353, 303, 354, 302, 354, 301, 353, 300, 353, 297,
                    352, 296, 352, 294, 351, 293, 352, 292, 351, 291, 351, 287,
                    350, 286, 350, 278, 349, 277, 349, 269, 348, 268, 348, 267,
                    347, 266, 347, 265, 345, 263, 345, 262, 344, 261, 343, 261,
                    332, 250, 331, 250, 330, 249, 329, 249, 328, 248, 327, 249,
                    326, 248
                ]
            ],
            "amodal_segmentation": [
                [
                    324, 248, 323, 249, 321, 249, 320, 250, 319, 250, 318, 251,
                    317, 251, 315, 253, 314, 253, 312, 255, 311, 255, 307, 259,
                    307, 261, 306, 262, 306, 264, 305, 265, 305, 276, 306, 277,
                    305, 278, 306, 279, 306, 283, 307, 284, 307, 285, 308, 286,
                    308, 289, 310, 291, 310, 292, 312, 294, 312, 296, 313, 297,
                    313, 298, 314, 299, 314, 300, 316, 302, 316, 303, 317, 304,
                    317, 305, 319, 307, 319, 308, 321, 310, 323, 310, 324, 311,
                    329, 311, 330, 312, 331, 312, 332, 313, 334, 313, 335, 314,
                    336, 314, 337, 313, 340, 313, 342, 311, 343, 311, 344, 310,
                    345, 310, 347, 308, 348, 308, 350, 306, 351, 306, 353, 304,
                    353, 303, 354, 302, 354, 301, 353, 300, 353, 297, 352, 296,
                    352, 294, 351, 293, 352, 292, 351, 291, 351, 287, 350, 286,
                    350, 278, 349, 277, 349, 269, 348, 268, 348, 267, 347, 266,
                    347, 265, 345, 263, 345, 262, 344, 261, 343, 261, 332, 250,
                    331, 250, 330, 249, 329, 249, 328, 248, 327, 249, 326, 248
                ]
            ]
        },

Data Generation & Pipeline

The creation of ACAD follows a hybrid Real-to-Synthetic (R2S) workflow, ensuring that the synthetic images retain highly realistic, real-world material textures.

1. The process began by collecting physical concrete aggregate samples. These real-world stones were photographed, manually annotated, and digitally extracted to build a high-quality source asset bank of individual aggregate templates with transparent backgrounds.
2. Using these real stone templates, the generation pipeline procedurally arranges them onto a textured concrete background. The system automatically resizes, rotates, and clusters the assets to simulate natural aggregate distribution under controlled low and high occlusion levels, applying consistent soft shadows for realistic depth.
3. For every placed object, the pipeline mathematically extracts both the Amodal Mask (the complete, occluded shape) and the Modal Mask (the visible portion). Finally, the images are rendered and automatically partitioned into standard training (80%), validation (10%), and testing (10%) splits.