Emerging Tracking from Video Diffusion
Keywords: Pixel-level object tracking, Temporal correspondence, Diffusion models
TL;DR: Our work achieves state-of-the-art performance in pixel-level object tracking by incorporating video diffusion model representations.
Abstract: We find video diffusion models, renowned for their generative capabilities, surprisingly excel at pixel-level object tracking without any explicit training for this task. We introduce a simple and effective method to extract motion representations from video diffusion models, achieving state-of-the-art tracking results. Our approach enables the tracking of identical objects, overcoming limitations of previous methods reliant on intra-frame appearance correspondence. Visualizations and empirical results show that our approach outperforms recent self-supervised tracking methods, including the state-of-the-art, by up to 6 points. Our work demonstrates video generative models can learn intrinsic temporal dynamics of video, and excel in tracking tasks beyond original video synthesis.
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: 278
Loading