Tracknet Tracking Eval

tracknet-tracking-eval

TrackNet: A Deep Learning Network for Tracking High-speed and Tiny Objects in Sports Applications — Huang et al. (2019) (arXiv:1907.03698, 2019)

What this evaluates

Evaluates the ability of deep learning models to detect and track high-speed, tiny objects (tennis and badminton balls) in broadcast sports videos. It probes robustness to motion blur, occlusion, and domain shifts by comparing single-frame vs. multi-frame tracking and transfer learning across different sports.

Datasets

• Tennis — total 20844; splits: train (-1), test (-1)
• Badminton — total 18242; splits: train (-1), test (-1)

Metrics

• Precision — range: percent
  • Precision = # of True Positive / (# of True Positive + False Positive).
• Recall — range: percent
  • Recall = # of True Positive / (# of VC1+VC2+VC3), where VC1-VC3 represent visible ball counts across visibility classes.
• F1-measure (primary) — range: percent
  • F1-measure = 2(Precision × Recall) / (Precision + Recall). Harmonic mean of precision and recall.

Input / output format

Input: Video frames resized to 640×360 pixels. Models take either a single frame or three consecutive frames as input.

Output: Heatmap indicating the predicted 2D coordinates of the tracked object (ball) in the target frame.

Scoring recipe

PE = euclidean_dist(pred_coords, gt_coords)
threshold = 5.0 if sport == 'tennis' else 7.5
if PE <= threshold: TP += 1
else: FP += 1
# FN counted from visibility classes (VC1-VC3)
precision = TP / (TP + FP)
recall = TP / (TP + FN)
f1 = 2 * (precision * recall) / (precision + recall)

Common pitfalls

• PE threshold is sport-dependent (5px for tennis, 7.5px for badminton), not a fixed global value.
• Recall denominator is defined by visible ball counts (VC1+VC2+VC3), not total frames, making it sensitive to occlusion/visibility classes.
• Transfer learning from tennis to badminton fails due to domain shift in speed and shape, not model architecture.

Evidence (verbatim from paper)

The overall performance in terms of precision, recall, and F1-measure are summarized in Table [V]. These three metrics are defined by Precision = # of True Positive / (# of True Positive + False Positive), Recall = # of True Positive / (# of VC1+VC2+VC3), and F1-measure = 2(Precision × Recall) / (Precision + Recall.

Citation

@misc{huang2019tracknet,
  title={TrackNet: A Deep Learning Network for Tracking High-speed and Tiny Objects in Sports Applications},
  author={Huang et al. (2019)},
  year={2019},
  note={arXiv:1907.03698}
}

• arXiv: 1907.03698