Skeleton-based interactive object co-part segmentation

Abstract

Object co-part segmentation, which involves segmenting shared objects into meaningful parts in a group of images, is a challenging joint-processing task. Although fully unsupervised deep learning algorithms exist for this task, the resultant parts often lack semantic meaning. This is because these algorithms use latent space to separate the parts, which may not necessarily correspond to meaningful parts as perceived by humans. Additionally, the number of parts required by these algorithms is difficult to pre-determine due to pose and size variations shared objects may exhibit across images, making human interaction necessary. While some interactive methods exist, none of them have explored the use of skeletons, which provide an object structure that can be leveraged to generate meaningful parts. Our proposed approach addresses this gap by presenting a skeleton-based interactive co-part segmentation framework that draws benefits from both unsupervised deep learning and human interaction. The framework employs the correspondence capabilities offered by deep learning counterparts and utilizes skeletons to generate meaningful parts. Experiments on Pascal-Part dataset demonstrate that our proposed framework outperforms existing interactive co-part segmentation methods in terms of segmentation accuracy and meaningfulness of parts.