--- license: apache-2.0 tags: - 3d-object-detection - multi-object-tracking - gaussian-splatting - camera-radar-fusion - nuscenes datasets: - nuscenes --- # GaussianMOT Joint 3D object detection and tracking on a camera–radar Gaussian BEV. Code: ## Files | File | Size | Content | |---|---|---| | `gaussianmot_epoch23.ckpt` | 368 MB | paper model, end of the 24th training epoch | | `gaussianmot_retrained_epoch23.ckpt` | 368 MB | paper model + 24 epochs with two training-label fixes (see the code README, *Post-submission update*) | | `fusion_warmstart.ckpt` | 288 MB | camera–radar Gaussian fusion checkpoint without the tracking head; `warm_start_path` for training | | `env/` | | build environment of these checkpoints: `conda_environment.yml`, `requirements_freeze.txt`, `versions.txt` (PyTorch 2.4.1 + CUDA 11.8, RTX 3090) | All checkpoints contain `state_dict` only (no optimizer state) and load with `tools/evaluate.py` as is. SHA-256: ``` d3e3a5e55265b878420dbbcf1d6868e47350e8c968fb5061e2aad3b88af5dd23 gaussianmot_epoch23.ckpt ad9e08e3bf5f13ee865c7e0f718a64a4b939b7a251c63c6bf2e73e95a8feb9b5 gaussianmot_retrained_epoch23.ckpt f9b28a22eb8e60a6c0b187753a57a2090f73d38e17963bc8afc8742ab81d1f4b fusion_warmstart.ckpt ``` ## Results (nuScenes val, inference seed 0) Paper model with the paper's evaluation settings (paper Tables I and II): | Detection | | Tracking | | |---|---|---|---| | mAP | 35.0 | AMOTA | 0.348 | | NDS | 42.5 | AMOTP | 1.31 | | mATE | 0.539 | Recall | 0.411 | | mASE | 0.307 | ID-S | 897 | | mAOE | 0.681 | | | | mAVE | 0.511 | | | | mAAE | 0.465 | | | 6 cameras + 5 radars, 448×800 input, 6.5 FPS on an RTX 3090 (batch size 1). Post-submission results (`attr_rule=standard`, which labels vehicles at or below 0.2 m/s as parked; *tuned* post-processing was selected on the validation split, see the code README): | Checkpoint | Post-proc. | mAP | NDS | mAOE | mAVE | AMOTA | AMOTP | Recall | ID-S | |---|---|---|---|---|---|---|---|---|---| | `gaussianmot_epoch23` | paper | 35.0 | 44.7 | 0.681 | 0.511 | 0.348 | 1.31 | 0.411 | 897 | | `gaussianmot_epoch23` | tuned | 38.0 | 46.3 | 0.677 | 0.507 | 0.389 | 1.23 | 0.462 | 1084 | | `gaussianmot_retrained_epoch23` | paper | 34.5 | 46.5 | 0.617 | 0.428 | 0.356 | 1.32 | 0.437 | 927 | | `gaussianmot_retrained_epoch23` | tuned | 37.1 | 48.0 | 0.607 | 0.424 | 0.397 | 1.25 | 0.454 | 980 | ## Usage ```bash git clone https://github.com/KimJunHan/GaussianMOT && cd GaussianMOT hf download kimjunhan/GaussianMOT --local-dir ckpt # paper numbers python tools/evaluate.py checkpoint_path=ckpt/gaussianmot_epoch23.ckpt device=cuda # retrained checkpoint, tuned post-processing python tools/evaluate.py checkpoint_path=ckpt/gaussianmot_retrained_epoch23.ckpt device=cuda \ attr_rule=standard \ 'decoder.nms_circle_radii=[1.8,3.0,4.5,3.0,3.0,0.75,0.8,0.8,0.3,1.0]' \ '+decoder.nms_rescale_factors=[1.0,0.7,0.4,0.55,0.7,1.5,1.0,1.0,3.5,1.1]' \ +tracker.init_score_thr=0.2 ``` Densification draws random shifts at inference, so results depend on `eval_seed` (default 0) in the last digit.