Table I — Dataset Scale and Coverage Summary
3,879
Number of annotations
51h 7m
Annotated-hours of reviewed scenes
24h 11m
Unique annotated driving time (overlap merged)
9,438
Equivalent doScene annotations (~19.5s each)
47.4s
Mean segment duration
31.5s
Median segment duration
15.0s
Minimum segment duration
254.4s
Maximum segment duration
35
Clips with zero annotations
Table II — Language Statistics
2,906
Number of instructions
18.25
Mean instruction length (words)
13.0
Median instruction length (words)
2
Minimum instruction length (words)
157
Maximum instruction length (words)
47.87%
Multi-sentence instructions
Table III — Annotation Diversity and Overlap
3.69
Mean annotations per clip
3.00
Median annotations per clip
13
Max annotations for a single clip
87.8%
Clips with multiple annotations
6,543
Overlapping annotation pairs
0.466
Mean temporal overlap ratio
0.255
Mean instruction similarity for overlaps
Figures
Click an annotator's bar in Fig 1 to filter the other figures to that annotator; click it again to reset.
Coverage Timeline (annotated vs. available, by city)
annotated downloaded, not annotated
Random Scene — Annotation Overlap
Data quality
Unrecognized commentary values (not in COMMENTARY_MAP): {'Ambiguous': 28, 'Static-Referential': 4, 'Dynamic-Referential': 2, 'BoStatic Referentialth': 1}
4308 duplicate row(s) dropped during merge.
2 file(s) failed to parse and were skipped.
35 clip(s) in the library have zero annotations so far. Full list written to unannotated_clips.csv alongside this dashboard.