GaussRender: Learning 3D Occupancy with Gaussian Rendering

Loïck Chambon    Éloi Zablocki    Alexandre Boulch    Mickaël Chen    Matthieu Cord

arXiv 2025

Paper   Code   

Scene visualization 1
Scene visualization 2
GaussRender is a 3D Occupancy module that can be plugged into any 3D Occupancy model to enhance its predictions and ensure 2D-3D consistency while improving mIoU, IoU, and RayIoU.

Abstract

Understanding the 3D geometry and semantics of driving scenes is critical for safe autonomous driving. Recent advances in 3D occupancy prediction have improved scene representation but often suffer from spatial inconsistencies, leading to floating artifacts and poor surface localization. Existing voxel-wise losses (e.g., cross-entropy) fail to enforce geometric coherence. In this paper, we propose GaussRender, a module that improves 3D occupancy learning by enforcing projective consistency. Our key idea is to project both predicted and ground-truth 3D occupancy into 2D camera views, where we apply supervision. Our method penalizes 3D configurations that produce inconsistent 2D projections, thereby enforcing a more coherent 3D structure. To achieve this efficiently, we leverage differentiable rendering with Gaussian splatting. GaussRender seamlessly integrates with existing architectures while maintaining efficiency and requiring no inference-time modifications. Extensive evaluations on multiple benchmarks (SurroundOcc-nuScenes, Occ3D-nuScenes, SSCBench-KITTI360) demonstrate that GaussRender significantly improves geometric fidelity across various 3D occupancy models (TPVFormer, SurroundOcc, Symphonies), achieving state-of-the-art results, particularly on surface-sensitive metrics.


GaussRender can be plugged to any model. The core idea is to transform voxels into gaussians before performing a depth and a semantic rendering.

Results

GaussRender can be plugged into any 3D model. We have dedicated experiments on multiple 3D benchmarks (SurroundOcc-nuScenes, Occ3D-nuScenes, SSCBench-KITTI360) and on multiple models (TPVFormer, SurroundOcc, Symphonies) to evaluate its performance.

Occ3D-nuScenes

Models TPVFormer (ours) TPVFormer SurroundOcc (ours) SurroundOcc OccFormer RenderOcc
Type w/ GaussRender base w/ GaussRender base base base
mIoU 30.48 🥇 (+2.65) 27.83 30.38 🥈 (+1.17) 29.21 21.93 26.11
RayIoU 38.3 🥇 (+1.1) 37.2 37.5 🥈 (+2.0) 35.5 - 19.5

SurroundOcc-nuScenes

Models TPVFormer (ours) TPVFormer SurroundOcc (ours) SurroundOcc OccFormer GaussianFormerv2
Type w/ GaussRender base w/ GaussRender base base base
IoU 32.05 🥈 (+1.19) 30.86 32.61 🥇 (+1.12) 31.49 31.39 30.56
mIoU 20.58 🥈 (+3.48) 17.10 20.82 🥇 (+0.52) 20.30 19.03 20.02

SSCBench-KITTI360

Models SurroundOcc (ours) SurroundOcc Symphonies (ours) Symphonies OccFormer MonoScene
Type w/ GaussRender base w/ GaussRender base base base
IoU 38.62 (+0.11) 38.51 44.08 🥇 (+0.68) 43.40 🥈 40.27 37.87
mIoU 13.34 (+0.26) 13.08 18.11 🥇 (+0.29) 17.82 🥈 13.81 12.31

Updates


BibTeX

@misc{chambon2025gaussrenderlearning3doccupancy,
      title={GaussRender: Learning 3D Occupancy with Gaussian Rendering}, 
      author={Loick Chambon and Eloi Zablocki and Alexandre Boulch and Mickael Chen and Matthieu Cord},
      year={2025},
      eprint={2502.05040},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.05040}, 
}