MCTR: Multi Camera Tracking Transformer

This repository provides training and evaluation code for MCTR using MMPTracking and MTMC_NVIDIA datasets.

Clone Repo

git clone <repo_address>
make pull # pull submodules

Create Env

mamba env create -f mcmot39
mamba activate mcmot39

Training

MCTR training is performed in 2 steps:

1. First step uses pairwise_init.yaml config
2. Second step uses pairwise.yaml

We used slurm for training jobs and corresponding scripts are available with the config files.

Evaluation

Evaluation script: scripts/trackeval_mmptrack.py & scripts/trackeval_trackbox_mmptrack.py performs the inference using the pretrained checkpoint. The trackeval_trackbox_mmptrack.py script uses the bounding box predictions from the trackbox head. It generates prediction and groundtruth files which are then used by trackeval library to run the MOT metrics. Cross camera metrics are output to the stdout when the script completes

Other

The scripts directory contains other useful sanity checks and visualization scripts that may be helpful for debugging.