Volleyball Group Activity Recognition — Baseline 8
Hierarchical temporal model for the Volleyball dataset (Ibrahim et al., CVPR 2016): a frozen ResNet-50 feature extractor per player crop → player LSTM over 9 frames → pooling separately per team (left/right, split by box x-position) → scene LSTM with skip connections → 8-way group-activity classifier.
Results (test split, 1,337 clips, 16 videos)
| Model | Accuracy | Macro-F1 |
|---|---|---|
| B1 single-frame ResNet-50 | 62.60% | 0.630 |
| B7 hierarchical, side-blind pooling | 73.75% | 0.701 |
| B8 (this model), team-split pooling | 85.64% | 0.855 |
| B9 YOLO26-x full-frame classifier, scored per clip | 80.33% | 0.817 |
(The paper reports 81.9% for its hierarchical model.)
Usage
Needs b8_model.py from the Space/repo. Input: crops (1, 9, P, 3, 224, 224) (ImageNet-normalised player crops),
masks (1, P), team_ids (1, P) (0 = left, 1 = right). Output: logits over
l-pass, r-pass, l-spike, r_spike, l_set, r_set, l_winpoint, r_winpoint.
Try it in the Space.
Limitations
- Requires tracked player boxes. There is no detector or tracker in the model; on new video you must supply them, and accuracy with predicted (rather than ground-truth) boxes has not been measured.
- Trained and evaluated on one dataset (broadcast-style volleyball, 55 videos); no claim of generalisation.
- Per-clip prediction from 9 frames centred on the annotated frame.