TUMTraf CenterPoint EP180 — 3rd Place @ DriveX 2026

🌟 Why This Model is Used & Its Benefits

• Benefit: Cameras can get blinded by sun glare, shadows, or dark colors. Because this model relies only on physical laser bounces, it is absolutely flawless at detecting hard, metallic objects.
• Our Result: This model achieved a perfect 1.0 (100%) score on Motorcycles, and we used it to surgically "fill in" missing Cars and Vans that the Camera model missed.

🛠️ Training Process & Data Pipeline

1. Data Processing

• Uses a MeanVFE (Voxel Feature Encoder) to compress 3D points inside each voxel into a single mean value.
• Compresses the 3D space down to a 2D Bird's Eye View (BEV) heatmap using HeightCompression with 384 BEV features.

2. Training Hyperparameters

• Detection Head (CenterHead with 3 separate class groupings):
  • What it is: We separated classes into: Group 1 (CAR, VAN), Group 2 (TRUCK, TRAILER, BUS), Group 3 (PEDESTRIAN, BICYCLE, MOTORCYCLE).
  • Why we chose this: If a tiny motorcycle and a massive bus share the exact same neural network detection head, the massive loss penalty from the bus overpowers the motorcycle. Separating them by size allows the model to specialize its weights for small vs. large objects.
• Optimizer (adam_onecycle with Peak LR 0.003):
  • Why we chose this: The 1-cycle learning rate schedule allows the model to train extremely fast and escape local minimums by ramping up the learning rate very high early in training, then slowly decaying it for fine-tuning at the end.
• Loss Weights (loc_weight: 2.0, cls_weight: 1.0):
  • Why we chose this: In autonomous driving, being 1 meter off in localization causes a crash, whereas being unsure if an object is a "Van" or a "Car" is less critical. We heavily penalized localization errors to force the model to draw incredibly tight, accurate bounding boxes.

3. How Accuracy Was Tracked

• Trained for 200 Epochs.
• We evaluated the model using the strict NMS_THRESH: 0.2 (Non-Maximum Suppression).
  • Why we chose this threshold: It aggressively removes duplicate overlapping boxes. V2X data is notorious for duplicate detections (e.g., the car sees the object, and the street pole sees the same object). A low threshold ensures only one box survives.
• The model achieved a flawless validation score on Motorcycles at Epoch 180, which perfectly translated to a 1.0 AP on the hidden CodaBench test set.

📂 Files

🏆 Results (CodaBench — Final Submission)