deepsworld/mctr-mmptrack

MCTR — Multi Camera Tracking Transformer (MMPTracking checkpoints)

Checkpoints that reproduce the MMPTrack validation results of

MCTR: Multi Camera Tracking Transformer, Alexandru Niculescu-Mizil, Deep Patel, Iain Melvin. arXiv:2408.13243 · code

MCTR is an end-to-end multi-camera multi-object tracker: a DETR-style detector per camera view, a shared set of track embeddings updated every frame, and soft probabilistic track↔detection association trained with differentiable losses. A single finetuned checkpoint per environment serves both reported variants:

• MCTR — outputs the per-view detection boxes (scripts/trackeval_mmptrack.py)
• MCTR-TB — outputs the track-head predicted boxes (scripts/trackeval_trackbox_mmptrack.py)

Each checkpoint is finetuned on one environment (fixed number of cameras / clips), so there is one model per scene: cafe, industry, lobby, office, retail.

Contents

Scene	Cameras	File	Size (GB)	SHA-256 (first 16)
cafe	4	cafe/mctr_cafe_epoch99.pth	0.23	069d9a899211d902
industry	4	industry/mctr_industry_epoch99.pth	0.23	7f59c4e85afadad6
lobby	4	lobby/mctr_lobby_epoch99.pth	0.23	4421777c62f7c076
office	5	office/mctr_office_epoch99.pth	0.25	36db4869e9246262
retail	6	retail/mctr_retail_epoch99.pth	0.26	f5ba93e1c4a5085f

Each scene folder contains:

• mctr_<scene>_epoch99.pth — the finetuned checkpoint (epoch 99 of a 100-epoch finetune of the 2-stage training protocol: pairwise_init.yaml → pairwise.yaml)
• train_config.yaml — exact training config of the run
• metrics.json — provenance + metrics (paper reference and reproduction)

The .pth is a torch.save dict with keys cfg (yacs config of the run), state_dict (the PAIRWISE model, already de-module.-prefixed), loss, epoch. It is loaded by main_pairwise._build_model in the training repo with strict=True; the camera count and clip set are fixed by the embedded config.

Usage

git clone https://github.com/necla-ml/mctr && cd mctr && make pull
mamba env create -f mcmot39        # conda env from the repo
mamba activate mcmot39             # conda activate mcmot39

# point the eval scripts at the MMPTracking dataset root you use, then:
python scripts/trackeval_mmptrack.py          /path/to/mctr_cafe_epoch99.pth   # MCTR
python scripts/trackeval_trackbox_mmptrack.py /path/to/mctr_cafe_epoch99.pth   # MCTR-TB

Notes:

• The eval scripts hardcode the dataset root /net/mlfs02/data/projects/shared/datasets/MMPTracking/ and the MMPTracking clip layout (per-scene folders, 64pm subsample); edit cfg.DATASET.ROOT in the scripts to match your copy of the data.
• The scripts write tracklet files under scripts/eval_outputs/ and call submodules/trackeval/scripts/run_mot_challenge.py (HOTA/CLEAR/Identity, no preproc).
• Inference is online, frame-by-frame, batch size 1 (keep_prob=0.9); ~233 MB model, roughly linear cost in the number of cameras.