Visual tracking with online structural similarity-based weighted multiple instance learning

This paper presents an online adaptive tracker, which employs a novel weighted multiple instance learning (WMIL) approach. In the proposed tracker, both positive and negative sample importances are integrated into an online learning mechanism for improving tracking performance in challenging environments. The sample importance is computed based on a new measure, i.e., structural similarity (SSIM), instead of using the Euclidean distance. Moreover, a novel bag probability function, which adopts both positive and negative weighted instance probabilities, is designed. Furthermore, a novel efficient weak classifier selection solution is developed for the proposed tracker. Qualitative and quantitative experiments on 30 challenging image sequences show that the novel tracking algorithm, i.e., SSIM-WMIL tracker, performs favorably against the MIL and WMIL counterparts as well as other 13 recently-proposed state-of-the-art trackers in terms of accuracy, robustness and efficiency. In addition, the negative sample importance can be used to enhance the multiple instance learning, and the SSIM-based approach is capable of improving the multiple instance learning performance for object tracking when compared to the Euclidean distance-based method. (C) 2019 Elsevier Inc. All rights reserved.