Learning to segment videos in HEVC compressed domain

High efficiency video coding (HEVC), as an up-to-date video coding standard, is becoming widely used, due to its preeminent performance. We argue in this paper that the new HEVC encoder can be utilized to provide efficient features for video segmentation. Thus, we propose a novel method, which learns to segment videos in HEVC compressed domain. Specifically, three features are extracted from HEVC bitstreams. Then, these features are integrated for video segmentation in compressed domain, through the logistic regression which is learnt from training data. This way, the computational complexity on video decoding can be saved. Experimental results show that our method is significantly superior to other state-of-the-art methods in terms of F-measure.