Multi-Object Tracking with Multiple Cues and Switcher-Aware Classification

[jiunbae]

Paper

• paper

• score(mota): 54.7

  References

• [ ] Online multi-object tracking with dual matching attention networks. (ECCV 2018)

• [ ] Tracking the un- trackable: Learning to track multiple cues with long-term dependencies. (ICCV 2017)

• [ ] On- line multi-object tracking with historical appearance match- ing and scene adaptive detection filtering. (arXiv 2018)

• [ ] Online multi-object tracking using cnn-based single ob- ject tracker with spatial-temporal attention mechanism. (ICCV 2017)

[jiunbae]

Abstract

unified MOT framework includes a SOT sub-net to capture short term, ReID sub-net to extract long term and a switcher-aware classifier to make matching decisions.

• short term cues find FN. Updated cues between neighbouring frames.
• long term cues avoid critical mistakes when occlusion. Stands for tracklet-long cues containing appearance features of object.
• SAC(switcher-aware classifier) learns to combine multiple cues

Introduction

The main contributions of paper:

1. Learning to capture long term and short term cues and making adaptive decisions for robust tracking
2. A switcher-aware classification(SAC) in data association for improving the robustness of MOT to identity switch.

Framework

• Base Tracker: Siamese-RPN tracker for extracting short term cues.
• ReID Network: Using Inception-v4 last FC layer before classification.
• Switcher-Aware Classifier