Tracktor[1] is used to detect objects from videos in each frame, while forming tracks by linking corresponding detections across time. There are some details of reading and implementing it.

Contents

• Paper & Code & note
• Paper understanding
• Code understanding
• Note
• References

Paper & Code & note

Paper: [Tracking without bells and whistles(ICCV 2019 paper)
Code: PyTorch
Note: tracking_wo_bnw

Paper

Abstract

As abstract of the paper, their work mainly converted a detector into a Tracktor, which exploit the bounding box regression of an object detector to predict the position of an object in the next frame.

1. It extended a straightforward re-identification and camera motion compensation to improving identity preservation across frames.
2. It achieved state-of-the-art on tackling most of the easy tracking scenarios. Besides, it also got ideal effect in tackling complex tracking scenarios. Therefore, it point out promising future research directions.

Problem Description

It shows the problem of multi-object tracking and exsiting solution for tacking this problem.

Problem Solution

It intrudued a Tracktor, converted detector into a tracktor by exploit the bounding box regression, and then extended Siamese network and motion model.

Conceptual Understanding

It describes the whole flow of Tracktor, including detector and two processing steps, that is initialing new tracks and killing old tracks. Then it explains each symbol of this process and how to deal with them.

Details of implementation

tracking multi-object

1. tracking step: it includes detecting object and form tracks by linking frames.
2. reID: it uses Siamese network to generate apearance feature to re-identify killed objects.
3. motion model: it contains problems of large camera motion and low video frame rate.

Experiments

1. tracking: it choosed Faster R-CNN with ResNet-101 as backbone network, it also provided FPN and other strategy.
2. reID: it trained TriNet with ResNet-50 as backbone network, and triplet loss as loss function.
3. CMC and ECC to deal with large camera motion, CVA to tackle low video frame rate problem.

Code

The complete code can be found in here with citing tracking_wo_bnw[2].

Tracktor algorithm

Note

some methods to improve accuracy or accelerate speed can add into this program.

References

[1] Bergmann, Philipp, Tim Meinhardt, and Laura Leal-Taixe. “Tracking without bells and whistles.” arXiv preprint arXiv:1903.05625 (2019).
[2] tracking_wo_bnw. https://github.com/phil-bergmann/tracking_wo_bnw