BFTracker: A One-Shot Baseline Model with Fusion Similarity Algorithm Towards Real-Time Multi-object Tracking

The key problem of real-time tracking is to achieve the balance between the tracking accuracy and real-time inference. Currently, the one-shot tracker, which joins multiple tasks into a single network, achieves trade-offs between tracking speed and accuracy. Different from previous trackers’ practice of exchanging more extra calculation cost for tracking accuracy, a new one-shot baseline that is simple and efficient is proposed by us. We continue to discuss the conflict under the tracking paradigm of joint detection and re-identification tasks and try to explore the source of this conflict and also devote to alleviate the feature conflict in the one-shot model. On the other hand, we propose a fusion similarity algorithm to focus on handling tracking challenges of sudden appearance changes and potential Kalman Filter prediction errors. On the MOT17 test set, the algorithm proposed can reduce the ID switches by 42.2% compared with the initial algorithm. On the MOT16 test set, the proposed BFTracker improves the ID F1 Score (i.e. IDF1) by 2.8% compared with the most popular one-shot tracker. In particular, Proposed baseline is quite simple and runs at 31 FPS on a single GPU.