CB-FQA: Custom Basketball Footwork Quality Assessment Dataset (Full Real-World Collection)

Release Notes for v2.0.0 (Important Clarification) This v2.0.0 release contains the complete 2,865 real-world standardized adversarial drill sequences captured via a multi-camera optical motion capture system. It entirely supersedes the preliminary v1.0 demo subset (which contained 500 sequences for early code verification). Data Collection & Composition Participants: 30 trained male collegiate athletes. Twelve defender-attacker pairs contributed to the training set. The validation and test sets each contain three defenders, with standardized attacker partners who are not among the 30 recorded athletes. Categories: Lateral Sliding (1,215), Crossover Transition (782), Close-out & Contesting (545), and Backpedal & Recovery (323). Sequence Generation Logic: Data was collected over a 4-week camp. Athletes performed 6-8 trials per category per session. Out of the 2,940 theoretical raw trials, strict occlusion and tracking failure screening yielded the final 2,865 high-fidelity sequences. Hardware & Signal Processing: Recorded at 60 fps with 12 synchronized cameras. Extracted 25-joint 3D spatial coordinates were smoothed using a temporal Gaussian kernel (tau=2, sigma=1.0). An average of 0.8% missing coordinates were interpolated using the cubic spline method. Quality Annotation Each sequence is annotated with a continuous quality score [0, 100]. Annotations are the arithmetic mean of independent evaluations from 5 senior, nationally certified basketball coaches. Evaluation criteria included center-of-gravity control, power continuity, and the effectiveness of spatial pressure. Directory Structure /Motion_Data: Contains .npy files structured as [T, 25, 3] for both defenders and attackers, alongside sequence-specific JSON metadata. /Quality_Scores: Contains JSON scorecards and a consolidated quality_scores_summary.csv for ICC computation. /Metadata: Includes the data dictionary, detailed experimental protocols, and demographic distributions. splits.json & participants.tsv: Strict subject-level Train/Validation/Test split mappings preventing data leakage.