CFSORT: improved SORT based on coarse to fine mechanism for object tracking

Abstract

In recent years, object tracking attracts much attention as a upstream task in many specific area. Multi-object tracking is one of them and its challenge are ID switch and ID re-matching due to object occlusion and image quality. In this paper, two possible methods are introduced to improve the baseline. In ReID model, a data augmentation method is built. Its main idea is to discuss and reduce the influence of background by retraining the ReID model. As ReID focuses on the features of objects in most cases, backgrounds can affect the feature metric significantly. In this situation, different objects in different backgrounds are captured and extracted as train data. Specially, a pure background, or background without objects are used in triplet loss to maximum the similarity distance. Besides, an updated SORT is implemented as CFSORT. In CFSORT, features are extracted by retrained ReID and two new processes or units are created. Comparing with the baseline, CFSORT requires lower confidence objects and emphasis a “coarse to fine” process. Those two new units are corresponding to similarity and IOU measurement. Currently, all new detection will be treated equally at first matching action. Similarly, more detection and tracked objects will match in new IOU unit which is stricter than the normal one. As a consequence, the result indicates that CFSORT has an improved performance on several evaluation scores in different scales. Meanwhile, this work can also prove the importance of accepting low confidence.