PROJECT TITLE :

Exploring Weakly Labeled Images for Video Object Segmentation With Submodular Proposal Selection - 2018

ABSTRACT:

Video object segmentation (VOS) is important for numerous computer vision issues, and handling it with minimal human supervision is very desired for the large-scale applications. To bring down the supervision, existing approaches largely follow a data mining perspective by assuming the supply of multiple videos sharing the identical object classes. It, however, would be problematic for the tasks that consume a single video. To address this downside, this Project proposes a unique approach that explores weakly labeled pictures to resolve video object segmentation. Given a video labeled with a target class, pictures labeled with the identical class are collected, from which noisy object exemplars are automatically discovered. When that the proposed approach extracts a collection of region proposals on various frames and efficiently matches them with large noisy exemplars in terms of look and spatial context. We then jointly select the most effective proposals across the video by solving a novel submodular problem that mixes region voting and global region matching. Finally, the localization results are leveraged as sturdy supervision to guide pixel-level segmentation. In depth experiments are conducted on 2 challenging public databases: Youtube-Objects and DAVIS. The results counsel that the proposed approach improves over previous weakly supervised/unsupervised approaches significantly, showing a performance even comparable with the many approaches supervised by the pricey manual segmentations.