SlowFast R50 Soccer Play-State Classifier
This binary video classifier predicts whether a soccer broadcast clip is in active play. It was locally fine-tuned from the PyTorchVideo SlowFast R50 model pretrained on Kinetics-400.
Model details
- Architecture: SlowFast R50
- Task: binary video classification
- Input: 32 uniformly sampled RGB frames at 224 × 224
- Classes:
0 = not_in_round,1 = in_round - Selected checkpoint: epoch 5
- Validation accuracy: 89.94%
- Validation loss: 0.4052
- Base model: PyTorchVideo SlowFast R50 pretrained on Kinetics-400
The reported metric comes from the local validation split and has not been independently reproduced on a public benchmark.
Intended use
Research and prototyping for soccer broadcast segmentation, active-play filtering, highlight extraction and editing assistance. The model is not an official event detector, identity system or substitute for human editorial judgment.
Training procedure
The PyTorchVideo backbone was frozen and the final classification block was fine-tuned using AdamW, cosine scheduling, label-smoothed cross-entropy and mixed precision.
- Epochs: 8
- Batch size: 4
- Learning rate:
5e-5 - Weight decay:
1e-5 - Clip length: 32
- Normalization: mean
[0.45, 0.45, 0.45], standard deviation[0.225, 0.225, 0.225]
Training-data disclosure
Training videos are not distributed. The local dataset contained soccer broadcast clips grouped into active and inactive play. Some source footage was collected from publicly accessible Bilibili videos, but public availability does not necessarily grant redistribution rights. No source video, frame, audio, subtitle, uploader information or platform metadata is included.
The source footage has not undergone complete work-by-work copyright clearance. Users must evaluate their intended use under applicable copyright, privacy, publicity and platform rules. Rights holders may request review or removal through the Hugging Face repository contact/discussion channel.
Usage
pip install -r requirements.txt
python inference.py path/to/video.mp4 --checkpoint model.pth
model.pth contains model_state_dict, epoch and validation metadata. Load pickle-based PyTorch checkpoints only from trusted sources.
Limitations
- Replays, graphics, crowd shots and unusual broadcasts may be misclassified.
- Generalization to non-broadcast or non-soccer video is not established.
- The model may learn broadcaster-specific visual cues.
- It has not been audited for memorization or demographic bias.
- Do not use it for surveillance or biometric identification.
License and attribution
Apache License 2.0. See LICENSE and NOTICE. The base PyTorchVideo project is Apache-2.0.