VisualComputingInstitute / towards-reid-tracking

Code for the paper "Towards a Principled Integration of Multi-Camera Re-Identification and Tracking through Optimal Bayes Filters"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Towards a Principled Integration of Multi-Camera Re-Identification and Tracking through Optimal Bayes Filters

This is the code for reproducing the experiments from our paper Towards a Principled Integration of Multi-Camera Re-Identification and Tracking through Optimal Bayes Filters. If you end up using any of this in your publication or otherwise find it useful, please cite our work as:

@article{BeyerBreuers2017Arxiv,
  author    = {Lucas Beyer and
               Stefan Breuers and
               Vitaly Kurin and
               Bastian Leibe},
  title     = {Towards a Principled Integration of Multi-Camera Re-Identification
               and Tracking through Optimal Bayes Filters},
  journal   = {arXiv preprint arXiv:1705.04608},
  year      = {2017},
}

Please note that this is very much research code, and the paper is a very exploratory one. It's made public for reference so that others can see what exactly we did, as the paper in no way can explain everything in enough detail. It is not production-quality code, rather it is nice code that got ever more messy as the deadline approached.

Due to the nature of the code, many things might still be confusing and non-obvious to others, so feel free to ask us, either by opening an issue here on github (preferably), or shooting us an e-mail!

The neural networks

The training code of the neural networks is not public yet as it's pending publication of the dependency at https://github.com/VisualComputingInstitute/triplet-reid.

However, the code creating the models and loading the trained weights is included. It is based on a custom deep-learning library on top of Theano called DeepFried2 and a small toolbox called lbtoolbox that you'll need to install. This can be easily done using pip install -e git+GITHUB_URL, see the corresponding READMEs.

The model we used for final experiments is lunet2c and the weights we used can be downloaded here.

The dataset and evaluation

This experimental work has been evaluated on the dukeMTMC dataset. Please refer to this project page for the used images, annotations, evaluation script, etc.

The run parameters

The below settings correspond to Table 1 of the paper. Details on the parameters can be found in Section 4.

NN-KF
DIST_THRESH = 200, det_init_thresh = 0.3, det_continue_thresh = 0.0 init_thresh = 3, delete_thresh = 5

+GT init
--gt_init
DIST_THRESH = 200, DET_INIT_THRESH = 0.3, DET_CONTINUE_THRESH = -0.3, init_thresh=1, delete_thresh=90

+ReID
--gt_init --use_appearance
DIST_THRESH = 200, APP_THRESH = 6, DET_INIT_THRESH = 0.3, DET_CONTINUE_THRESH = -0.3, init_thresh=1, delete_thresh=90

only ReID
--gt_init --use_appearance
DIST_THRESH = 6, DET_INIT_THRESH = 0.3, DET_CONTINUE_THRESH = -0.3, init_thresh=1, delete_thresh=90

Full
--dist_thresh 6 --unmiss_thresh 2

+entropy
--dist_thresh 5.5 --ent_thresh 0.25 --maxlife 8000 --unmiss_thresh 5
killed of age: 4

Final raw bounding box results can be found here.

About

Code for the paper "Towards a Principled Integration of Multi-Camera Re-Identification and Tracking through Optimal Bayes Filters"

License:MIT License


Languages

Language:Python 100.0%