Answers are grounded in this paper and cite their sources.
In recent years, joint detection and embedding (JDE) has become the research focus in multi-object tracking (MOT) due to its fast inference speed. JDE models are designed and widely utilized to train the detection task and the re-identification (Re-ID) task jointly. However, there exists a severe issue overlooked by previous JDE models, i.e., the detection task requires category-level features but the Re-ID task requires instance-level features. This could lead to feature conflict, which would hurt the performance of JDE models. Furthermore, inaccurate detection results can degrade the final tracking accuracy even when discriminative Re-ID features are provided. In this article, we propose a new balancing method for training JDE models, which monitors the training process of the detection task and adjusts the weights of the detection task and Re-ID task in the training phase. Our proposed balancing method ensures a well-trained detection model and a good trade-off between the detection task and Re-ID task. Comprehensive experiments on two public MOT benchmarks demonstrate the effectiveness and superiority of our proposed balancing method. In particular, our proposed balancing method could achieve new state-of-the-art results on MOT challenges without additional training data.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2022) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: