Title:Efficient Object Tracking based on Circular and Structural Multi-level Learners

Abstract:We propose a novel efficient tracking framework. Firstly, we incorporate DCF and SOSVM to obtain a novel formulation for training circular and structural learners (CSL). Secondly, we introduce a collaborative optimization strategy to update the learners, which significantly reduces computational complexity and improves robustness. Thirdly, we observe the fact that features extracted from only single-level are not robust to handle all challenge factors, thus we suggest to get a multi-level confidence score map with deep features in the continuous spatial domain, and we exploit an implicit interpolation model to extract multi-resolution complementary deep features based on different pre-trained CNNs, including both the deep appearance features and deep motion features of the target. Finally, in order to get an optimal confidence score map for more accurate localization, we propose a novel ensemble post-processor which based on relative entropy to combine the sing-level confidence score maps. Comprehensive evaluations are conducted on three object tracking benchmarks. Our approach obtains an absolute gain of 0.3% and 0.6% in mean AUC score compare to the top-ranked method on the OTB-2013 and OTB-2015 benchmarks, respectively, and provides a third-best performance with expected average overlap (EAO) score of 29.8% on the VOT2017 challenge, while operates at frame-rate.