Candidate Selection-based Deep Affinity Network for Multi-object Tracking

Deep Affinity Network (DAN) is a novel approach in multi-object tracking (MOT) designed to jointly modeling object appearances and affinities end to end. But tracking accuracy of DAN tracker is greatly limited since it neglects unreliable detection. Exploiting predictions of tracks has emerged as a popular approach to tackle the task of tracking-by-detection. However, it's observed that missing detection has not been solved well enough which would significantly influence tracking accuracy. Thus, obtaining more reliable tracking candidates is concerned to further address the problem of missing detection. In this paper, we propose Candidate Selection-based Deep Affinity Network (CSDAN) tracker for MOT. It collects candidates from detection, predictions of tracks and backward tracking simultaneously so that they can complement each other in different scenarios. Moreover, we propose a deep learned candidate selection model (DCSM) with a unified scoring function suitable for CSDAN, which can well handle candidates from three sources separately and select those for data association. Experiments conducted on MOT17 benchmark demonstrate that our extensions can significantly address the unreliable detection problem in DAN tracker, and our CSDAN tracker demonstrates competitive tracking performance.