Robust object recognition in RGB-D egocentric videos based on Sparse Affine Hull Kernel

In this paper, we propose a novel kernel function for recognizing objects in RGB-D egocentric videos. In order to effectively exploit the varied object appearance in a video, we take a set-based recognition approach and represent the target object using the set of frames contained in the video. Our kernel function measures the similarity of two sets by the minimum distance between the sparse affine hulls of the two sets. Our kernel function also allows convenient integration of heterogeneous data modalities beyond RGB and depth. We extensively evaluate the proposed method on three benchmark datasets, including two RGB-D object datasets and one thermal/visible face dataset. All the results clearly show that the proposed method outperforms state-of-the-art methods.