Research map / 3D vision / 3D generation

QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction

NeurIPS 2025 · 3D vision › 3D generation · approach: Vision-based 3D occupancy prediction · 2 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 QuadricFormer, which represents driving scenes with sparse superquadric primitives (instead of ellipsoidal Gaussians) and a probabilistic superquadric mixture for vision-centric 3D semantic occupancy prediction on nuScenes.

Problem

Efficient and expressive 3D semantic occupancy prediction from multi-view images; dense voxel methods are costly and Gaussian object-centric methods need many ellipsoidal primitives to model diverse shapes.

Approach

Use superquadrics (with shape exponents epsilon1, epsilon2) as object-centric scene primitives, model each as a local occupancy probability distribution with geometry prior, aggregate via probabilistic mixture, and add a pruning-and-splitting module that concentrates primitives in occupied regions.

Key results

• QuadricFormer reaches 20.12 mIoU and 31.22 IoU on nuScenes with 1600 superquadrics, exceeding GaussianFormer-2 (20.02 mIoU, 30.56 IoU) which uses 12800 Gaussians.
• At 1600 primitives QuadricFormer runs at 162 ms latency and 2554 MB memory on an RTX 4090, versus 341 ms and 3075 MB for depth-initialized GaussianFormer-2 at the same count.
• Best ablation setting of the superquadric exponent range (0.1, 2) gives 20.51 mIoU and 31.25 IoU at 12800 superquadrics.

Measured effects reported in the paper

Limitations

• Random initialization means superquadrics cannot fully learn accurate positions, leaving some in empty regions and reducing efficiency.
• Evaluated only on nuScenes with dense occupancy annotations from SurroundOcc.

Experimental setup

Datasets: nuScenes SurroundOcc dense occupancy annotations

Benchmarks: nuScenes occupancy validation (SurroundOcc annotations)

Metrics: mIoU IoU latency (ms) memory (MB)

Baselines: MonoScene Atlas BEVFormer TPVFormer OccFormer SurroundOcc GaussianFormer GaussianFormer-2

Builds on: GaussianFormer-2 GaussianFormer Lovasz-softmax loss FCOS3D

Closest papers by meaning

• GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D Occupancy Prediction (CVPR 2025 · similarity 0.92)
• ODG: Occupancy Prediction Using Dual Gaussians (NeurIPS 2025 · similarity 0.90)
• Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.90)
• OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.89)
• GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction (CVPR 2025 · similarity 0.88)
• CSV-Occ: Fusing Multi-frame Alignment for Occupancy Prediction with Temporal Cross State Space Model and Central Voting Mechanism (ICML 2025 · similarity 0.88)