Improved MDNet tracking with fast feature extraction and efficient multiple domain training

Haoyue Shi, Jifeng Ning, Yangchen Fu, Jing Ni

Abstract: MDNet tracking method draws great attention due to its high precision and robustness on several evaluation datasets. However, its pretrain method for feature representation and speed of feature extraction still have some limitations. In this work, we present an improved MDNet tracking algorithm to overcome the above problems. First, we propose an efficient multiple domain mechanism to train more robust deep feature. For each domain, we employ multiple classification simultaneously to identify different interested targets from multiple videos instead of original binary classification only to identify the target and background from one sequence. Second, the proposed method accelerates feature extraction procedure by using RoIAlign layer based on VGG-M network. Our algorithm, denoted by FE-MDNet, which means fast feature extraction and efficient multiple domain training for MDNet, is evaluated on OTB2015 and TrackingNet. The results show that our algorithm performs 16 times faster than MDNet with better accuracy compared to MDNet and demonstrates favorably against state-of-the-art tracking methods.