2019/12/03 by Sungkwon Choo, Choo, Sungkwon, Wonkyo Seo +3
Computer Science · #Advanced Image and Video Retrieval Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Video Surveillance and Tracking Methods #Visual Attention and Saliency Detection
paper · pdf · doi:10.48550/arxiv.1912.01373
openalex publication_date 2019/12/03 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
This paper presents a method for automatic video object segmentation based on\nthe fusion of motion stream, appearance stream, and instance-aware\nsegmentation. The proposed scheme consists of a two-stream fusion network and\nan instance segmentation network. The two-stream fusion network again consists\nof motion and appearance stream networks, which extract long-term temporal and\nspatial information, respectively. Unlike the existing two-stream fusion\nmethods, the proposed fusion network blends the two streams at the original\nresolution for obtaining accurate segmentation boundary. We develop a recurrent\nbidirectional multiscale structure with skip connection for the stream fusion\nnetwork to extract long-term temporal information. Also, the multiscale\nstructure enables to obtain the original resolution features at the end of the\nnetwork. As a result of two-stream fusion, we have a pixel-level probabilistic\nsegmentation map, which has higher values at the pixels belonging to the\nforeground object. By combining the probability of foreground map and\nobjectness score of instance segmentation mask, we finally obtain foreground\nsegmentation results for video sequences without any user intervention, i.e.,\nwe achieve successful automatic video segmentation. The proposed structure\nshows a state-of-the-art performance for automatic video object segmentation\ntask, and also achieves near semi-supervised performance.\n