FairMOT: On the fairness of detection and re-identification in multi-target tracking
Summary:
In recent years, significant progress has been made in object tracking and re-identification, which is a key component for multi-target tracking. However, few people pay attention to jointly accomplishing these two tasks in a single network. Our research shows that previous attempts eventually led to a decrease in accuracy, mainly because the task of re-identification was not learned fairly, which resulted in a very pair of identity switching. This unfairness exists in two aspects: (1) They regard the accuracy to a large extent on the re-id of the first detection task as the second task, so the training is heavily biased towards the detection task but ignores the re-id task (2) They use ROI-Align to extract re-id features directly borrowed from object detection. However, this introduces a lot of ambiguity when characterizing objects, because many sampling points may belong to interference instances or backgrounds. To solve this problem, we propose a simple method FairMOT, which consists of two homogeneous branches to predict the pixel-level target score and re-identify features. The fairness achieved between tasks enables FairMOT to obtain a high level of detection and tracking accuracy, and has a greater advantage over the previous technical level on several public data sets.