Research map / 3D vision / 3D generation

Outdoor Scene Extrapolation with Hierarchical Generative Cellular Automata

CVPR 2024 · 3D vision › 3D generation · approach: Generative indoor scene synthesis · 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 hierarchical Generative Cellular Automata (hGCA), a coarse-to-fine conditional 3D generative model with a BEV planner that extrapolates high-resolution outdoor street geometry from sparse LiDAR scans, trained on synthetic CARLA and Karton City data.

Problem

Generating fine-grained, simulation-ready 3D street geometry beyond occlusions and the field of view from sparse, noisy LiDAR scans of autonomous vehicles, where prior semantic scene completion methods give low-resolution outputs and local GCA lacks global context.

Approach

Two-stage coarse-to-fine generation: a GCA conditioned on features from a light-weight bird's-eye-view planner produces low-resolution voxels with global consistency, then a continuous GCA (cGCA) upsamples them to high-resolution implicit surfaces.

Key results

• On CARLA with 5 scans, hGCA (10cm implicit) reaches avg High LiDAR ReSim 3.25/2.99 range improvements over GCA (avg 5.83 vs ~4.65 for hGCA implicit), outperforming all baselines on reconstruction metrics while keeping diversity (TMD).
• On Karton City with 5 scans, hGCA implicit achieves High LiDAR ReSim avg 3.20 versus 3.95 for GCA at 20cm and 4.29 for cGCA.
• On real Waymo scans, hGCA generates more complete geometry qualitatively, though its quantitative LiDAR ReSim/IoU is slightly worse than SCPNet and SG-NN on held-out real scans.

Measured effects reported in the paper

Surprising findings

• The planner trades off diversity for quality and global consistency, lowering TMD relative to vanilla GCA.
• Quantitative scores on real Waymo data were slightly worse than baselines even though qualitative completeness was better, showing the difficulty of evaluating on real data.

Limitations

• Generative process is slow and not real-time.
• Generated roofs and geometry on real data are sometimes inconsistent; fidelity inferior to synthetic data.
• Quantitative evaluation on real data is difficult; real-data metrics are slightly worse than baselines.

Experimental setup

Datasets: CARLA Karton City (Turbosquid asset) ShapeNet car assets Waymo Open Dataset 3DFront (pretrained autoencoder) Objaverse-XL (out-of-distribution asset)

Benchmarks: CARLA Karton City Waymo Open Dataset

Metrics: High LiDAR ReSim (chamfer distance) IoU at 20cm voxels Street CD Total Mutual Difference (TMD)

Baselines: JS3CNet SCPNet SG-NN ConvOcc GCA cGCA

Builds on: Generative Cellular Automata cGCA PointPillars SPADE infusion training

Closest papers by meaning

• CymbaDiff: Structured Spatial Diffusion for Sketch-based 3D Semantic Urban Scene Generation (NeurIPS 2025 · similarity 0.88)
• GenAssets: Generating in-the-wild 3D Assets in Latent Space (CVPR 2025 · similarity 0.87)
• Towards Realistic Scene Generation with LiDAR Diffusion Models (CVPR 2024 · similarity 0.87)
• StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models (CVPR 2025 · similarity 0.87)
• Genesis: Multimodal Driving Scene Generation with Spatio-Temporal and Cross-Modal Consistency (NeurIPS 2025 · similarity 0.86)
• LidaRF: Delving into Lidar for Neural Radiance Field on Street Scenes (CVPR 2024 · similarity 0.86)