Traffic Accident Understanding on ACCIDENT Leaderboard (private)
59.94COurs 32B (3-stage)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Ours 32B (3-stage)Backbone=Qwen3-VL-32B, Stages=3-stage2026.05 | 59.94 | 56.37 | 56.03 | 53.5 | 56.74 | |
| Ours 9:1 ensemble (32B type)Backbone=32B/235B blend, Ensemble ratio=9:12026.05 | 59.94 | 56.57 | 56 | 54.1 | 56.948 | |
| Ours + bbox snap (final)Backbone=32B/235B blend, Post-processing=bbox snap, Ensemble ratio=9:12026.05 | 59.94 | 56.69 | 56 | 54.41 | 57.08 | |
| Baseline AType=Rule-based, Time=mid-clip, Spatial=frame center2026.06 | 58.07 | — | 18.96 | 25.05 | 27.34 | |
| Baseline BType=Rule-based, Time=quarter-clip, Spatial=frame center2026.06 | 58.07 | — | 26.64 | 25.05 | 31.88 | |
| Ours 235B (3-stage)Backbone=235B MoE, Stages=3-stage2026.05 | 57.03 | 55.04 | 53.88 | 54.3 | 55.67 | |
| Three-stage pipelineTemporal=PE δ-window, Classification=entropy-gated pairwise, Spatial=OWL-v2 type+scene conditioned2026.06 | 50.57 | — | 36.89 | 34.98 | 40.15 | |
| Molmo-7B baselineBackbone=Molmo-7B2026.05 | 29.3 | 35.8 | 34.3 | 48.8 | — |