Adaptive Fusion of RGBD Data for two-stream FCN-based Level Set Tracking
Jie Yang, Xue Song Zhou, Zheng Ou Zhou, Hao Wen · 2019
In this paper, we address the contour tracking task based on a deep fully convolutional network (FCN)-driven Level Set method. Different from previous Level Set-based trackers wherein speed functions are usually modeled by handcrafted features. In our work by considering the merits of FCN on its powerful feature representation and end-to-end pixelwise prediction capabilities, we aim to integrate it into Level Set contour evolution framework. Specifically, we import RGBD data and design a two-stream FCN-based structure. An Adaptive fusion mechanism is applied to leverage these two complementary modalities. We adopt the final fusion map from the two-stream FCN output as a guidance to evolve the contour more efficiently and accurately. Experimental results on a number of video clips demonstrate the effectiveness and robustness of our proposed method, especially for tackling the challenging problems, i.e., similar appearance between target and background, and low illumination, etc.