harphool17/tumtraf-ensemble-pipeline

TUMTraf V2X Ensemble Pipeline — 3rd Place @ DriveX 2026

🌟 Why This Pipeline is Used & Its Benefits

No single AI model is perfect. Some models are great at seeing colors (Cameras), while others are great at seeing shapes in the dark (LiDAR).

• Benefit: By creating an Ensemble (a team of models), we get the "best of both worlds". If Model A misses a car because it's painted black in the shadows, Model B will catch it using laser distance.

🧠 Glossary for Beginners

• Ensemble: Combining multiple AI models together to vote on the final answer, resulting in a much higher score than any single model could get alone.
• Surgical Non-Overlap Fill: A smart coding trick we used. We take high-confidence predictions from our LiDAR model, but only place them into empty spaces where our Camera model didn't see anything. This prevents duplicate boxes.
• Multi-Epoch Union: Instead of picking the single best training checkpoint, we take the predictions from 7 different stages of the AI's training and merge them together. This drastically improved our ability to find small, fast-moving Pedestrians.

⚙️ Ensemble Strategy

1. Per-class best model selection: Hardcoded the pipeline to choose the historically best model for each specific vehicle type.
2. Multi-epoch PEDESTRIAN union: Merged predictions from epochs 46, 48, 50, 52, 55, 58, and 60 (score>0.30, dist>1.5m).
3. Surgical non-overlap fill: CenterPoint Ep150 adds missed boxes for VAN/TRAILER/CAR.

🏆 Final Results — 0.9062 mAP (3rd Place)

Class	Precision	Recall	AP	Model
BUS	1.000	1.000	1.000	CoopDet3D base
MOTORCYCLE	1.000	1.000	1.000	CenterPoint Ep180
VAN	1.000	0.950	0.950	CoopDet3D ped-boost Ep18 + CP fill
TRUCK	1.000	0.941	0.941	CoopDet3D ped-boost Ep18
TRAILER	0.994	0.931	0.925	CoopDet3D ped-boost Ep18 + CP fill
CAR	0.967	0.902	0.872	CoopDet3D Ep60 + CP fill
BICYCLE	0.952	0.909	0.866	CoopDet3D Ep60
PEDESTRIAN	0.822	0.847	0.696	Multi-epoch union (Ep46-60)
Overall			0.9062