Skip to content

README.md

Latest commit

 

History

History
133 lines (102 loc) · 4.62 KB

README.md

File metadata and controls

133 lines (102 loc) · 4.62 KB

RacketPose

Racket detection and 5-keypoint pose estimation module, built on RTMDet (detection) and RTMPose (pose estimation) from the OpenMMLab ecosystem.

Architecture

The module runs a two-stage pipeline:

  1. RTMDet-M detects racket bounding boxes (3 classes: badminton / tabletennis / tennis racket).
  2. RTMPose-M estimates 5 keypoints per detected racket: top, bottom, handle, left, right.

Directory Structure

RacketPose/
├── configs/
│   ├── detection/
│   │   ├── rtmdet_m_racket.py          # Full training config
│   │   ├── rtmdet_m_racket_infer.py    # Inference-only config (no _base_)
│   │   └── rtmdet_toy.py              # Toy config for pipeline testing
│   ├── pose/
│   │   ├── rtmpose_m_racket.py         # Full training config
│   │   ├── rtmpose_m_racket_infer.py   # Inference-only config (no _base_)
│   │   └── rtmpose_toy.py             # Toy config for pipeline testing
│   └── _base_/                         # Shared base configs
├── checkpoints/
│   ├── epoch_300.pth                   # RTMDet pre-trained weights
│   └── best_PCK_epoch_90.pth          # RTMPose pre-trained weights
├── datasets/                           # Custom dataset classes
├── evaluation/                         # Custom metrics
├── tools/
    │   ├── train_detection.py             # RTMDet training entry point
    │   ├── train_pose.py                  # RTMPose training entry point
    │   ├── test_detection.py              # RTMDet evaluation entry point
    │   ├── test_pose.py                   # RTMPose evaluation entry point
    │   └── inference.py                   # Batch det+pose inference
└── README.md

Environment

RacketPose requires mmdet and mmpose (tested with mmengine==0.10.7, mmcv==2.1.0, mmdet==3.3.0, mmpose==1.3.2). Use the same uball conda environment as the rest of the project; see the repository README for installation instructions.

Training

Detection (RTMDet)

conda activate uball
python tools/train_detection.py configs/detection/rtmdet_m_racket.py

Pose Estimation (RTMPose)

python tools/train_pose.py configs/pose/rtmpose_m_racket.py

Both scripts use data_root from the config's _base_/datasets/ files. The default expects COCO-format annotations in data/info/.

Evaluation

Detection (RTMDet)

python tools/test_detection.py configs/detection/rtmdet_m_racket.py \
       --checkpoint checkpoints/epoch_300.pth

Pose Estimation (RTMPose)

python tools/test_pose.py configs/pose/rtmpose_m_racket.py \
       --checkpoint checkpoints/best_PCK_epoch_90.pth

Both test scripts load the checkpoint, disable training components, and run runner.test(). You can override config values with --cfg-options, e.g.:

python tools/test_detection.py configs/detection/rtmdet_m_racket.py \
       --checkpoint checkpoints/epoch_300.pth \
       --cfg-options test_dataloader.batch_size=16

Inference

Run detection + pose estimation on all clips in a split:

python tools/inference.py \
    --det_config configs/detection/rtmdet_m_racket_infer.py \
    --det_ckpt checkpoints/epoch_300.pth \
    --pose_config configs/pose/rtmpose_m_racket_infer.py \
    --pose_ckpt checkpoints/best_PCK_epoch_90.pth \
    --data_root ../data \
    --sport badminton \
    --split test \
    --device cuda

Output: ../data/<sport>/pred_racket/<match>/<rally>/result.json

Each output JSON is a list of frames. Each frame contains a list of detected rackets:

[
  [
    {
      "bbox": [[x1, y1, x2, y2]],
      "bbox_score": 0.95,
      "keypoints": [[x1, y1], [x2, y2], [x3, y3], [x4, y4], [x5, y5]],
      "keypoint_scores": [0.9, 0.8, 0.7, 0.6, 0.5]
    }
  ],
  [],
  ...
]

Inference Options

Flag Default Description
--sport badminton Target sport (badminton / tabletennis / tennis)
--bbox_thr 0.3 Detection confidence threshold
--device cuda cuda or cpu
--debug off Process only the first clip

Config Notes

  • Training configs (rtmdet_m_racket.py, rtmpose_m_racket.py) use _base_ inheritance and require dataset registration via import datasets.
  • Inference configs (*_infer.py) are standalone single-file configs with no _base_ dependencies. They use BN instead of SyncBN for compatibility with single-GPU and CPU inference.
  • Detection uses num_classes=3 (one per sport's racket type), not 4.