A Study on Efficient Multi-task Networks for Multiple Object Tracking

Multiple object tracking involves multi-task learning to handle object detection and data association tasks concurrently. Conventionally, object detection consists of object classification and object localization (e.g., object regression) tasks, and data association is treated as a classification task. However, various tasks can cause inconsistent learning due to that the learning targets of object detection and data association tasks are different. Object detection focuses on positional information of objects while data association requires strong semantic information to identify same object target. Besides, advantageous character of multi-task learning is the correlation between tasks, and adopting such character in learning the networks can result in better generalization performance. However, existing multiple object tracking methods learn this information by treating multi-task branches independently. To understand the behaviours of multi-task networks in multiple object tracking, in this paper, we explore task-dependent representations through empirical experiments and observe that multi-task branches in multiple object tracking are complementary. To better learn such information, we introduce a novel Correlation Estimation (CE) module to estimate the correlation between object classification and bounding box regression based on statistical features of box regression quality. Finally, extensive experiments are conducted on the benchmark dataset MOT17. As a result, our method outperforms state-of-the-art online trackers without requiring additional training datasets.