Research map / 3D vision / 3D generation

GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction

CVPR 2025 · 3D vision › 3D generation · approach: Vision-based 3D occupancy prediction · 18 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GaussianFormer-2, a probabilistic Gaussian superposition model for 3D semantic occupancy prediction that treats each Gaussian as an occupancy probability, aggregates geometry via probabilistic multiplication, uses a Gaussian mixture for semantics, and initializes Gaussians with a distribution-based module.

Problem

3D semantic occupancy prediction from camera images is computationally redundant with dense grids, and existing 3D semantic Gaussian representations waste many Gaussians on empty space and suffer from overlap.

Approach

Interpret each Gaussian as a probability that its neighborhood is occupied, aggregate geometry via the multiplication theorem of probability (1 - prod(1 - alpha)), model semantics as a Gaussian mixture posterior with softmax-normalized classes, and initialize Gaussians by supervising pixel-aligned occupancy distributions along camera rays instead of LiDAR depth.

Key results

• On nuScenes surround-view occupancy, Ours (Ch.=192) reaches 31.74 IoU and 20.82 mIoU, above GaussianFormer's 29.83 IoU and 19.10 mIoU, using 12800 Gaussians versus 144000.
• On SSCBench-KITTI-360 monocular occupancy, the method reaches 13.90 mIoU, surpassing GaussianFormer's 12.92 mIoU by about 7.6% relative.
• With 25600 Gaussians the method achieves 20.33 mIoU at 3063 MB memory, versus GaussianFormer's 16.00 mIoU at 4850 MB with 25600 Gaussians.

Measured effects reported in the paper

Surprising findings

• The GaussianFormer-2 latency is higher than GaussianFormer's (e.g., 357 ms vs 227 ms at 25600 Gaussians), attributed to farthest point sampling in the initialization module.
• Using LiDAR point cloud initialization yields higher mIoU (21.17) than the proposed occupancy-distribution initialization (20.32) in the ablation, though it requires LiDAR supervision.

Limitations

• Initialization module adds latency due to farthest point sampling.
• Monocular KITTI-360 results are mixed: IoU is 38.37 versus 40.27 for OccFormer and some classes remain weaker.

Experimental setup

Datasets: nuScenes SurroundOcc occupancy annotations KITTI-360 SSCBench-KITTI-360

Benchmarks: nuScenes SSCBench-KITTI-360

Metrics: IoU mIoU latency memory

Baselines: GaussianFormer MonoScene BEVFormer TPVFormer OccFormer SurroundOcc Atlas Voxformer LMSCNet SSCNet

Builds on: GaussianFormer BEVFormer SurroundOcc TPVFormer

Closest papers by meaning

• ODG: Occupancy Prediction Using Dual Gaussians (NeurIPS 2025 · similarity 0.92)
• GaussTR: Foundation Model-Aligned Gaussian Transformer for Self-Supervised 3D Spatial Understanding (CVPR 2025 · similarity 0.92)
• GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction (CVPR 2025 · similarity 0.92)
• QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction (NeurIPS 2025 · similarity 0.92)
• Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.89)
• OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.88)