Correlation Filters with Weighted Convolution Responses

Zhiqun He, Yingruo Fan, Junfei Zhuang, Yuan Dong, Hongliang Bai

Abstract: In recent years, discriminative correlation filters based trackers have shown dominant results for visual object tracking. Combining the online learning efficiency of the correlation filters with the discriminative power of CNN features has aroused great attention. In this paper, we derive a continuous convolution operator based tracker which fully exploits the discriminative power in the CNN feature representations. In our work, we normalize each individual feature extracted from different layers of the deep pre-trained CNN first, and after that, the weighted convolution responses from each feature block are summed to produce the final confidence score. By this weighted sum operation, the empirical evaluations demonstrate clear improvements by our proposed tracker based on the Efficient Convolution Operators Tracker (ECO). On the other hand, we find the 10-layers design is optimal for continuous scale estimation, which contribute most to the performance. Finally, our tracker ranks top among the state-of-the-art trackers on VOT2016 dataset and outperforms the ECO tracker on VOT2017 dataset.