TransTrack: Multiple-Object Tracking with Transformer

  • 2020-12-31 06:03:00
  • Peize Sun, Yi Jiang, Rufeng Zhang, Enze Xie, Jinkun Cao, Xinting Hu, Tao Kong, Zehuan Yuan, Changhu Wang, Ping Luo
  • 35

Abstract

Multiple-object tracking(MOT) is mostly dominated by complex and multi-steptracking-by-detection algorithm, which performs object detection, featureextraction and temporal association, separately. Query-key mechanism insingle-object tracking(SOT), which tracks the object of the current frame byobject feature of the previous frame, has great potential to set up a simplejoint-detection-and-tracking MOT paradigm. Nonetheless, the query-key method isseldom studied due to its inability to detect new-coming objects. In this work,we propose TransTrack, a baseline for MOT with Transformer. It takes advantageof query-key mechanism and introduces a set of learned object queries into thepipeline to enable detecting new-coming objects. TransTrack has three mainadvantages: (1) It is an online joint-detection-and-tracking pipeline based onquery-key mechanism. Complex and multi-step components in the previous methodsare simplified. (2) It is a brand new architecture based on Transformer. Thelearned object query detects objects in the current frame. The object featurequery from the previous frame associates those current objects with theprevious ones. (3) For the first time, we demonstrate a much simple andeffective method based on query-key mechanism and Transformer architecturecould achieve competitive 65.8\% MOTA on the MOT17 challenge dataset. We hopeTransTrack can provide a new perspective for multiple-object tracking. The codeis available at: \url{https://github.com/PeizeSun/TransTrack}.

 

Quick Read (beta)

loading the full paper ...