dhvazquez/mtg_synthetic_cards_semantic_segmentation

Magic The Gatering synthetic image dataset.

Synthetic images of MTG cards with the following features:

• Photorealistic rendering using BlenderProc2 and HDRI environments
• Precise card geometry with rounded corners
• Random transformations for data augmentation
• Segmentation masks for semantic segmentation training
• Train
  • 6000 rendered images of normal cards (blender)
  • 2000 rendered images of Borderless cards (blender)
  • 800 Random Background images whit no cards (Lorem Picsum)
• Test
  • 6000 rendered images of normal cards (blender)
  • 2000 rendered images of Borderless cards (blender)
  • 180 Random Background images whit no cards (Lorem Picsum)

Dataset Structure

Train Set

/train/image            - Rendered Images (jpg)
/train/mask             - Mask (png)
/train/temp_textures    - 3d Model textures (reference)

Test Set

/test/image             - Rendered Images (jpg)
/test/mask              - Mask (png)
/test/temp_textures     - 3d Model textures (reference)

Corner annotations in corner_annotations.json

{
  "train": {
    "full_art_3a3afd00-da06-4a9f-8cd1-7133728e0fdd_002.jpg": [
      [
        137.5807150595883,
        185.1069880823402
      ],
      [
        335.0,
        181.81666666666666
      ],
      [
        335.0,
        441.74719101123594
      ],
      [
        146.23711810305517,
        451.29138166894666
      ]
    ],
    "normal_0bec6181-e214-4833-8c2f-8a10d59b2879_002.jpg": ...
  

Examples:

Image / Mask Examples:

Texture Example:
Corners Annotations Example:

For more details see github repo