b3d-fasterrcnn

A detectron2 Faster R-CNN X101-32x8d-FPN, fine-tuned from its COCO checkpoint on the human-labeled vehicle crops of the B3D dataset (Berkeley DeepDrive Drone, one vehicle class) at native 4K scale. It generated the pre-labels that were human-corrected into the B3D test-set ground truth used in the Tetris revision.

• config.yaml: the detectron2 config of the training run (input resizing disabled, 500 detections per image).
• model_final.pth: the final weights (sha256 f7caab8f1994c737eb3a3c5ac84a31a70db1b1651bff4ae43ceb66804d011174).

Training, detection, and tracking code: https://github.com/tetris-db/b3d-fasterrcnn (download_weights.sh fetches these files, and detect.py runs them).