BiC-Occ: Bi-directional Circulated 3D Occupancy Prediction for Autonomous Driving
Keywords: 3D occupancy prediciton, Bi-directional, View transformation, Autonomous driving
Abstract: Vision-based 3D occupancy prediction is the cornerstone in autonomous driving systems to provide comprehensive scene perception for subsequent decisions, which requires assessing voxelized 3D scenes with multi-view 2D images. Existing methods mainly adopt unidirectional pipelines projecting image features to BEV representations for following supervision, whose performances are limited by the sparsity and ambiguity of voxel labels. To address this issue, we propose a Bi-directional Circulated 3D Occupancy Prediction (BiC-Occ) framework for more accurate voxel predictions and supervisions. Specifically, we design a Bi-directional View Transformer module that approximates invertible transition matrices of the view transformation process, promoting the self-consistency between 2D image features and 3D BEV representations. Furthermore, we propose a Circulated Interpolation Predictor module that exploits local geometric structures to align multi-scale BEV representations, correcting local ambiguity with consistent occupancy predictions across different resolutions. With the synergy of these two modules, the self-consistency within different perception views and occupancy resolutions compensates for the sparsity and ambiguity of voxel labels, leading to more accurate 3D occupancy predictions. Extensive experiments and analyses demonstrate the effectiveness of our BiC-Occ framework.
Primary Area: applications to robotics, autonomy, planning
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: 658
Loading