Introduction

This repository contains the code for the paper "Can Deep Learning be Applied to Model-Based Multi-Object Tracking?" (https://arxiv.org/abs/2202.07909). The code for MT3v2 was developed as joint effort by Juliano Pinto, Georg Hess, and William Ljungbergh, and was partially based on the code available at the repositories for DETR and Deformable-DETR.

Setting up

In order to set up a conda environment with all the necessary dependencies, run the command:

 conda env create -f conda-env/environment-<gpu/cpu>.yml

Running experiments

Run an experiment using the training.py script. Example usage:

src/training.py -tp configs/tasks/task1.yaml -mp configs/models/mt3v2.yaml

Training hyperparameters such as batch size, learning rate, checkpoint interval, etc, are found in the file configs/models/mt3v2.yaml.

Evaluating experiments

You can plot metrics of interest using the util/plot_results.py script, during and after training.

Trained models

Trained models for each of the 4 tasks described in the paper can be found here.

The ablations trained for task 4 are available here.