spot_dat — BSL Fingerspelling Detector

Binary detector for fingerspelling segments in British Sign Language (BSL) video, trained on BOBSL.

Models

V3 Detector (models/variant3_75_25.pt)

1D CNN binary classifier over per-frame landmark features.

• Input: 158-dim features (hand landmarks + wrist distances + motion deltas) over 25-frame windows
• Architecture: 1D CNN, hard-negative trained (variant3). The 75/25 ratio is the hard/easy split within the negative class — 92,640 hard / 30,881 easy of 123,521 negatives — not a positive/negative split. See TRAINING_AUDIT_REPORT.md §6.2.
• Performance: 85.8% precision / 92.3% recall on BOBSL validation

Detection pipeline thresholds:

Parameter	Value	Description
t	0.87	Frame-level confidence threshold
k	11	Smoothing kernel size
gap	0.5s	Merge gap between adjacent detections
min_dur	0.5s	Minimum segment duration

GBT Filter (models/gbt_v3.pkl)

Gradient-boosted tree post-filter for removing false-positive detections.

• Input: 15 segment-level features (duration, confidence stats, hand presence, motion)
• Performance: CV F1 = 0.850
• Threshold: conf = 0.60

Scripts

Training & Evaluation

Script	Description
train_variants.py	Original variant training (pos/neg ratio sweep)
train_variants_v2.py	V2 training with hard negatives
extract_unified_dataset.py	Build training dataset from landmarks
extract_easy_neg_only.py	Extract easy negatives for baseline
mine_hard_negatives.py	Mine hard negatives from false positives
evaluate_v2.py	Evaluate detector variants on BOBSL
evaluate_bobsl.py	BOBSL-specific evaluation
evaluate_youtube.py	YouTube domain evaluation
verify_domain_shift.py	Domain shift analysis (v1)
verify_domain_shift_v2.py	Domain shift analysis (v2)
analyse_results.py	Result analysis and metrics
generate_report_v2.py	Generate evaluation reports
audit_negatives.py	Audit negative samples
plot_training_curves.py	Plot training loss/accuracy curves

Detection Pipeline (V3 + GBT)

Script	Description
frame_features.py	Extract 158-dim frame features from landmarks
run_full_detection_v3.py	Run V3 detector on full episodes
extract_segment_features.py	Extract 15-dim segment features for GBT
apply_gbt_filter.py	Apply GBT post-filter to detections
analyse_full_results_v3.py	Analyse full detection results
run_evaluation.py	End-to-end evaluation deep dive