Generalizable Dynamic Radiance Field in Egocentric View

DI QI, Tong Yang, Beining Wang, Xiangyu Zhang, Wenqiang Zhang

Keywords: generalized dynamic view synthesis, NeRF, computer vision

Abstract:

We present a novel framework for generalizable dynamic radiance field in egocentric view. Our approach can predict a 3D representation of the physical world at a given time based on a monocular video without test-time training. To this end, we use a contracted triplane as the 3D representation of physical world in an egocentric view at a specific time. To update the explicit 3D representation, we propose a 4D-aware transformer module to aggregate features from monocular videos. Besides, we also introduce a temporal-based 3D constraint to achieve better multiview consistency. In addition, we train the proposed model with large-scale monocular videos in a self-supervised manner. Our model achieves top results in novel view synthesis on dynamic scene datasets, demonstrating its strong understanding of 4D physical world. Besides, our model also shows the superior generalizability to unseen scenarios. Furthermore, we find that our approach emerges capabilities for geometry and semantic learning. We hope our approach can provide preliminary understanding of the physical world in first-person view and help ease future research in computer vision, computer graphics and robotics.

Primary Area: other topics in machine learning (i.e., none of the above)

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.

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: 5647