GlobalTrack[1] is a pure global tracker for long-term tracking, without temporal consistency assumption making cumulative errors. There are some details of reading and implementing it.

Contents

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

Paper & Code & note

Paper: GlobalTrack: A Simple and Strong Baseline for Long-term Tracking(AAAI 2020 paper)
Code: Pytorch
Note: GlobalTrack

Paper

Abstract

As abstract of the paper, their work mainly proposed a method called GlobalTrack, which is a pure global instance search based tracker that makes no assumption on the temporal consistency.

1. It is developed based on two-stage object detector Faster R-CNN, with two submodules QG-RPN and QG-RCNN.
2. it is able to perform full-image and multi-scale search of arbitrary instances with only a single query as the guide.
3. They further propose a cross-query loss to improve the robustness of this approach against distractors.

Problem Description

It shows the difficults of long-term tracking and the problem of existing trackers.

Problem Solution

It shows the methods for solving long-term tracking problem.

Conceptual Understanding

It describes the overall architecture of GlobalTrack with QG-RPN and QG-RCNN.

Details of implementation

1. Offline Training: it samples frame pairs from training videos.
2. Online Tracking: it contains initialization, tracking and results.
3. Cross-query Loss: it choose top-1 prediction as result.

Architecture

1. Query-Guide RPN: it generating query-specific proposals.
2. Query-Guide RCNN: it consists of feature modulation and traditional RCNN.
3. Tracking Results: it takes top-1 prediction as results.

Experiments

They compared this approach GlobalTrack with state-of-the-art trackers on four large-scale tracking benchmarks as follows.

Code

The complete code can be found in [GlobalTrack][2].

Note

some free ideas that orienting future work.

References

[1] Huang, Lianghua, Xin Zhao, and Kaiqi Huang. “GlobalTrack: A Simple and Strong Baseline for Long-term Tracking.” arXiv preprint arXiv:1912.08531 (2019).
[2] GlobalTrack. https://github.com/huanglianghua/GlobalTrack