Appearance Guidance Attention for Multi-Object Tracking

Abstract

Appearance information is one of the most important matching indicators for multi-object data association. In tracking by detection model, appearance information and detection information are usually integrated in the same sub-network for learning and output. This phenomenon will result in the appearance embedding vectors to be coupled to the network inference method during learning stage. As a result, the appearance embedding vectors contains too much background information and affects the accuracy of data association. Based on the non-end-to-end tracking model, we design an appearance guidance attention module for the appearance extraction branch. This module can effectively strengthen the network's learning of the object visual appearance features and reduce the attention in the learning of background features. Finally, we utilize the appearance embedding vectors that decoupled from the inference method as the input of the back-end tracker and perform data association. The proposed method is tested on the MOT16 and MOT17 datasets. Experiments show that the proposed method provides more high-quality appearance representation information for the back-end tracker and the tracking performance on two datasets is better than other comparison models. At the same time, our model can reach 24.9FPS on a single 1080Ti GPU.

Multiple object tracking using space-time adaptive correlation tracking

“…Filtering of detections is done using maximum suppression with a cost definition dependent on track and detection confidence. Chen et al [19] also employ a trained Siamese network for historical appearancebased matching in addition to motion and size-based matching, which special aids in lowering identity shifts during tracking are encountered. To calculate the total pair-wise cost between each detection and track, Song et al [20] uses two deep networks: the spatial attention network (which uses a siamese architecture to compare detection-bounding boxes and track history) and the temporal attention network (which uses an-longsort term memory (LSTM) architecture).…”