data/
├── train/          # Training set
│   ├── images/     # Training images
│   └── annotations/ # Training annotations (JSON format)
├── val/            # Validation set
│   ├── images/     # Validation images
│   └── annotations/ # Validation annotations (JSON format)
├── test/           # Test set
│   ├── images/     # Test images
│   └── annotations/ # Test annotations (JSON format)
└── external/       # External dataset (optional)
    └── external/
        ├── images/
        └── annotations/

Data Format

Image Format

Supported formats: JPEG (.jpeg, .jpg), PNG (.png)
Color space: RGB (configured via DATASET.COLOR_RGB in config)
Input size: Images will be resized and padded to 384×1152 (width×height)

Annotation Format

Annotations are stored in LabelMe JSON format. Each annotation file corresponds to one image file with the same base name.

JSON Structure

{
  "version": "2.4.4",
  "flags": {},
  "shapes": [
    {
      "label": "R_FC",
      "points": [[x1, y1], [x2, y2]],  // For circle: center and edge point
      "shape_type": "circle",           // or "point"
      ...
    },
    ...
  ],
  "imagePath": "image_filename.jpeg",
  "imageHeight": 8025,
  "imageWidth": 2947
}

Keypoint Types

Point: Single point annotation (shape_type: "point")
- points: [[x, y]] - keypoint coordinates
Circle: Circle annotation (shape_type: "circle")
- points: [[cx, cy], [edge_x, edge_y]] - center and edge point
- Note: Only the center point (points[0]) is used for keypoint detection

Keypoint Labels

The model expects 36 keypoints in a specific order. Keypoint labels include:

Hip region: R_FC, L_FC, R_GT, L_GT, R_FNeck_Cut_Up, L_FNeck_Cut_Up, R_FNeck_Cut_Down, L_FNeck_Cut_Down
Femur region: R_Cdy_Up, L_Cdy_Up, R_Cdy_Down, L_Cdy_Down
Knee region: Various knee-related keypoints
Tibia region: R_Cyd_Up, L_Cyd_Up, R_Cyd_Down, L_Cyd_Down
Ankle region: R_DLP, L_DLP, R_DMP, L_DMP, etc.

Data Preparation

1. Organize Your Data

Place your images and annotations in the corresponding directories:

Training images → data/train/images/
Training annotations → data/train/annotations/
Validation images → data/val/images/
Validation annotations → data/val/annotations/

2. File Naming Convention

Image and annotation files should have the same base name
Example:
- Image: 1_2_410_200049_2_47176438882121_3_1_20180304081116139_77313.jpeg
- Annotation: 1_2_410_200049_2_47176438882121_3_1_20180304081116139_77313.json

3. Annotation Requirements

Each annotation file must contain exactly 36 keypoints
Keypoints can be annotated as point or circle shapes
For circles, only the center point will be used
Missing keypoints should be padded with [0.0, 0.0] during evaluation

4. Image Preprocessing

The dataset loader automatically:

Resizes images to fit within 384×1152 while maintaining aspect ratio
Pads images to exactly 384×1152 (centered)
Converts grayscale images to 3-channel RGB
Applies color space conversion (BGR→RGB) if configured

Configuration

Set the following parameters in your config file (configs/llrformer.yaml):

DATASET:
  DATASET: 'MyKeypointDataset'  # Dataset class name
  ROOT: 'data'                   # Root directory
  TRAIN_SET: 'train'             # Training set folder
  TEST_SET: 'val'                # Validation set folder
  DATA_FORMAT: jpeg              # Image format (jpeg, png, or auto)
  COLOR_RGB: true                 # Use RGB color order

Dataset Statistics

Training set: 709 images
Validation set: 102 images
Test set: 200 images
External set: 136 images (optional)

Notes

Image Size: Original images can be any size. They will be automatically resized and padded to 384×1152 during loading.
Keypoint Order: The order of keypoints in the annotation file matters. Ensure consistency across all annotations.
Missing Keypoints: If a keypoint is missing in an annotation, it should be represented as [0.0, 0.0] or omitted (will be padded during evaluation).
External Dataset: The external/ folder contains additional data that may not follow the same structure. Adjust paths accordingly if using this data.
Format Support: The dataset loader supports multiple image formats. Set DATA_FORMAT: auto in config to automatically detect JPEG and PNG files.

Troubleshooting

Issue: Images not loading

Check file paths and naming conventions
Verify image format is supported (JPEG/PNG)
Ensure annotation files exist for each image

Issue: Keypoint count mismatch

Verify all annotations contain exactly 36 keypoints
Check for missing or extra keypoints in annotation files
Ensure keypoint order is consistent

Issue: Coordinate errors

Verify coordinate system (origin at top-left)
Check that coordinates are within image boundaries
Ensure annotation format matches LabelMe JSON structure

Downloads last month: 7