This repository provides the implementation of our paper Tracking without bells and whistles (Philipp Bergmann, Tim Meinhardt, Laura Leal-Taixe) [https://arxiv.org/abs/1903.05625]. This branch includes an updated version of Tracktor for PyTorch 1.3 with an improved object detector. The original results of the paper were produced with the iccv_19
branch.
In addition to our supplementary document, we provide an illustrative web-video-collection. The collection includes examplary Tracktor tracking results and multiple video examples to accompany our analysis of state-of-the-art tracking methods.
- Clone and enter this repository:
git clone https://github.com/phil-bergmann/tracking_wo_bnw
cd tracking_wo_bnw
-
Install packages for Python 3.7 in virtualenv:
pip3 install -r requirements.txt
- Install Tracktor:
pip3 install -e .
-
MOTChallenge data:
- Download MOT17Det, MOT16Labels, 2DMOT2015, MOT16-det-dpm-raw and MOT17Labels and place them in the
data
folder. As the images are the same for MOT17Det, MOT17 and MOT16 we only need one set of images for all three benchmarks. - Unzip all the data by executing:
unzip -d MOT17Det MOT17Det.zip unzip -d MOT16Labels MOT16Labels.zip unzip -d 2DMOT2015 2DMOT2015.zip unzip -d MOT16-det-dpm-raw MOT16-det-dpm-raw.zip unzip -d MOT17Labels MOT17Labels.zip
- Download MOT17Det, MOT16Labels, 2DMOT2015, MOT16-det-dpm-raw and MOT17Labels and place them in the
-
Download object detector and re-identifiaction Siamese network weights and MOTChallenge result files:
- Download zip file from here.
- Extract in
output
directory.
In order to configure, organize, log and reproduce our computational experiments we structured our code with the Sacred framework. For a detailed explanation of the Sacred interface please read its documentation.
-
Tracktor can be configured by changing the corresponding
experiments/cfgs/tracktor.yaml
config file. The default configuration runs Tracktor with the FPN object detector as described in the paper. -
The default configuration is
Tracktor
. RunTracktor
by executing:
python experiments/scripts/test_tracktor.py
- The results are logged in the corresponding
output
directory.
For reproducability, we provide the new result metrics of this updated code base on the MOT17
challenge. It should be noted, that these surpass the original Tracktor results. This is due to the newly trained object detector. This version of Tracktor does not differ conceptually from the original ICCV 2019 version (see branch iccv_19
). The results on the offical MOTChallenge webpage are denoted as the Tracktor v2
tracker. The train and test results are:
********************* MOT17 TRAIN Results *********************
IDF1 IDP IDR| Rcll Prcn FAR| GT MT PT ML| FP FN IDs FM| MOTA MOTP MOTAL
65.2 83.8 53.3| 63.1 99.2 0.11| 1638 550 714 374| 1732124291 903 1258| 62.3 89.6 62.6
********************* MOT17 TEST Results *********************
IDF1 IDP IDR| Rcll Prcn FAR| GT MT PT ML| FP FN IDs FM| MOTA MOTP MOTAL
55.1 73.6 44.1| 58.3 97.4 0.50| 2355 498 1026 831| 8866235449 1987 3763| 56.3 78.8 56.7
For the object detector we followed the new native torchvision
implementations of Faster-RCNN with FPN which are pretrained on COCO. The provided object detection model was trained and tested with this Google Colab notebook. The MOT17Det
train and test results are:
********************* MOT17Det TRAIN Results ***********
Average Precision: 0.9090
Rcll Prcn| FAR GT TP FP FN| MODA MODP
97.9 93.8| 0.81 66393 64989 4330 1404| 91.4 87.4
********************* MOT17Det TEST Results ***********
Average Precision: 0.8150
Rcll Prcn| FAR GT TP FP FN| MODA MODP
86.5 88.3| 2.23 114564 99132 13184 15432| 75.0 78.3
-
The training config file is located at
experiments/cfgs/reid.yaml
. -
Start training by executing:
python experiments/scripts/train_reid.py
If you use this software in your research, please cite our publication:
@InProceedings{tracktor_2019_ICCV,
author = {Bergmann, Philipp and Meinhardt, Tim and Leal{-}Taix{\'{e}}, Laura},
title = {Tracking Without Bells and Whistles},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}}