Grounding-PanopticSports Benchmark

Download the benchmark from here.

File Structure

The data is organized hierarchically as follows:

sequence_dir/
├── prompt_dir/
│   ├── cam0.pt
│   ├── cam1.pt
│   └── ...
└── ...

For example:

tennis/
├── tennis_shot/
│   ├── cam0.pt
│   ├── cam1.pt
│   └── ...
└── ...

• The sequence directory name (e.g., "tennis") is the data sequence
• The prompt directory name (e.g., "tennis_shot") is the action prompt

Data Format

Each .pt file is a PyTorch serialized dictionary containing segmentation masks and bounding boxes for a specific camera view, organized by the frame numbers in which the prompt occurs.

Frames without detections are not included in the dictionaries

Dictionary Structure

{
    "masks": {
        frame_number: segmentation_mask,
        ...
    },
    "boxes": {
        frame_number: bounding_box,
        ...
    }
}

• "masks": A dictionary where keys are frame numbers and values are segmentation masks for the prompted action
• "boxes": A dictionary where keys are frame numbers and values are bounding boxes for the prompted action

Example

For a file tennis/tennis_shot/cam0.pt, the structure might be:

{
    "masks": {
        10: [mask_data],
        11: [mask_data],
        ...
    },
    "boxes": {
        10: [x1, y1, x2, y2],
        11: [x1, y1, x2, y2],
        ...
    }
}