Sga Interact Eval

sga-interact-eval

SGA-INTERACT: A 3D Skeleton-based Benchmark for Group Activity Understanding in Modern Basketball Tactic — Yang et al. (2025) (arXiv:2503.06522, 2025)

What this evaluates

Evaluates models on group activity recognition (GAR) and temporal group activity localization (TGAL) using 3D skeleton sequences from basketball games. It probes spatio-temporal interaction modeling, long-term dependency handling, and the ability to leverage multi-view motion capture data for complex team tactics.

Datasets

• SGA-INTERACT — total ?; splits: train (-1), val (-1), test (-1); repo https://github.com/Charrrrrlie/SGA-INTERACT

Metrics

• accuracy (mAcc./Top3-mAcc./oAcc.) (primary) — range: [0, 1]
  • Mean accuracy (mAcc.) is the average of per-class accuracies. Top3-mAcc. measures the fraction of samples where the ground truth class appears in the top-3 predictions. Overall accuracy (oAcc.) is the global classification accuracy across all samples.
• mAP — range: [0, 1]
  • Mean Average Precision (mAP) for temporal boundary localization. Computed by calculating the Intersection over Union (IoU) between predicted and ground truth temporal intervals, applying confidence thresholds, and averaging the Area Under the Precision-Recall curve across all activity classes.

Input / output format

Input: 3D skeleton sequences (multi-view MoCAP data) representing player movements over time. Optionally augmented with ball possession and team information. For 2D baselines, sequences are projected into front and side views.

Output: For GAR: a single predicted group activity class label. For TGAL: temporal start and end boundaries (timestamps) for each group activity instance in untrimmed sequences.

Scoring recipe

# GAR Classification
pred_labels = np.argmax(model_output, axis=-1)
correct = (pred_labels == gold_labels)
mAcc = np.mean([correct[gold_labels == c].mean() for c in unique_classes])
top3_mAcc = np.mean([gold_labels[i] in np.argsort(model_output[i])[-3:] for i in range(N)])
oAcc = correct.mean()

# TGAL Localization
ious = compute_iou(pred_intervals, gt_intervals)
# Apply confidence threshold, compute precision-recall curve per class
ap_per_class = [compute_ap(precision, recall) for class in classes]
mAP = np.mean(ap_per_class)

Common pitfalls

• Reimplementing closed-source baselines (e.g., COMPOSER*, MPGCN*) may introduce subtle implementation differences affecting reproducibility.
• Applying combined spatial and temporal augmentations can degrade performance for certain models (e.g., MPGCN, One2Many-ARG) rather than improve it.
• Standard skeleton-based methods lack temporal dimension preservation, making them inapplicable to the TGAL task without architectural modifications.
• 2D projected inputs suffer from severe viewpoint sensitivity, leading to poor cross-view generalization compared to native 3D skeletons.

Evidence (verbatim from paper)

As shown in [Sec. 4.2], we evaluate all baselines on the GAR task using several metrics: accuracy, number of model parameters, and floating point operations (FLOPs). ... As shown in [Sec. 5.3], the proposed One2Many-STAtt demonstrates consistent superiority in the TGAL task. However, all baselines exhibit poor mAP performance.

Citation

@misc{yang2025sgainteract,
  title={SGA-INTERACT: A 3D Skeleton-based Benchmark for Group Activity Understanding in Modern Basketball Tactic},
  author={Yang et al. (2025)},
  year={2025},
  note={arXiv:2503.06522}
}

• arXiv: 2503.06522