$\text{O}_\text{2}$VIS: Occupancy-aware Object Association for Temporally Consistent Video Instance Segmentation
Keywords: Video instance segmentation, Long-term memory, Temprorally consistent learning
TL;DR: We propose an occupancy-guided memory and temporally consistent object association pipeline.
Abstract: In this paper, we present Occupancy-aware Object Association for Video Instance Segmentation ($\text{O}_{\text{2}}$VIS), a new framework crafted to improve long-term consistency in instance tracking. We introduce the Instance Occupancy Memory (IOM) that tracks global instance features and their occupancy status to effectively differentiate between recurring and new objects. It ensures consistent tracking and effective management of object identities across frames, enhancing the overall performance and reliability of the VIS process. Moreover, we propose a Decoupled Object Association (DOA) strategy that handles existing and newly appeared objects separately to optimally assign indices based on occupancy. This technique enhances the accuracy of object matching and ensures stable and consistent object alignment across frames, especially useful in dynamic settings where objects frequently appear and disappear. Extensive testing and an ablation study confirm the superiority of our method over traditional methods, establishing new standards in the VIS domain.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 638
Loading