Robust visual tracking by embedding combination and weighted-gradient optimization. (August 2020)

• Record Type: Journal Article
• Authors: Feng, Jin; Xu, Peng; Pu, Shi; Zhao, Kaili; Zhang, Honggang
• Is Part Of: Pattern recognition. Volume 104(2020:Aug.)
• Publication Date: 2020-08
• Subjects: Visual tracking -- Data imbalance -- Embedding combination -- Weighted-gradient loss; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107339
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a negative sample embedding combination network specific for handling the imbalance between positive and negative samples in the tracking-by-detection framework. We propose a weighted-gradient loss to handle the imbalance between easy and hard samples by balancing the total gradient contributions of them. We conduct extensive experiments on tracking benchmarks. The results demonstrate that the proposed algorithm improves the performance of the baseline and performs favorably against state-of-the-art trackers. Abstract: Existing tracking-by-detection approaches build trackers on binary classifiers. Despite achieving state-of-the-art performance on tracking benchmarks, these trackers pay limited attention to data imbalance issue, e.g, positive and negative, easy and hard. In this paper, we demonstrate that separately learning feature embeddings corresponding to negative samples with different semantic characteristics is effective in reducing the background diversity to handle the imbalance between positive and negative samples, which facilitates background awareness of classifiers. Specifically, we propose a negative sample embedding combination network, which helps to learn several sub-embeddings and combine them to build a robust classifier. In addition, we propose a weighted-gradient loss to handle the imbalance between easy and hard samples. The gradient contribution of each sample to model training is dynamically weighted according to the gradient distribution, which prevents easy samples from overwhelming model training. Extensive experiments on benchmarks demonstrate that our tracker performs favorably against state-of-the-art algorithms.