Reproduce “Height-Guided Projection Reparameterization for Camera-LiDAR Occupancy” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuan Wu, Zhiqiang Yan, Jiawei Lian, Zhengxue Wang, Jian Yang
HiPR (Height-Guided Projection Reparameterization) is a camera-LiDAR occupancy framework that dynamically reparameterizes the 2D-to-3D view transformation using LiDAR height maps to improve geometric alignment and feature aggregation. It achieves state-of-the-art results on Occ3D (54.7 mIoU, 50.0 RayIoU) and SurroundOcc (30.4 mIoU), with a lightweight variant maintaining real-time inference at 20.8 FPS.
Key insights
- Dynamically reparameterizing the 2D-to-3D projection space using a height prior, as opposed to fixed uniform sampling, significantly improves geometric alignment and feature aggregation for 3D occupancy prediction.
- Restricting feature aggregation to geometrically meaningful regions through a height-validity mask is crucial for suppressing noise from empty spaces and enhancing semantic consistency in BEV features.
- A progressive conditioning approach, gradually shifting from ground-truth supervision to noisy sensor data, is effective for stabilizing training and improving model robustness when leveraging imperfect sensor-derived priors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.