Research map / 3D vision / 3D generation

AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks

NeurIPS 2026 · 3D vision › 3D generation · approach: Vision-based 3D occupancy prediction · 0 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 AdaOcc, a point-based adaptive 3D semantic occupancy framework with a geometry-guided dual-branch encoder, progressive query learning for budget-adaptive inference, and a containment loss (C.P.O.), reaching SOTA on Occ-ScanNet.

Problem

Existing occupancy prediction methods do not adapt to varying compute budgets, sensor setups (depth or LiDAR inputs), and numbers of views, which limits deployment in embodied systems.

Approach

Predict sparse semantic points from geometry-initialized queries with a multi-layer progressive decoder whose query count and depth set the compute budget, plus a containment loss that penalizes predicted points lying in supervised empty space near occupied voxels.

Key results

• On Occ-ScanNet, AdaOcc reaches 59.67 mIoU and 65.29 IoU, improving over GPOcc-VGGT (56.19 mIoU, 63.14 IoU) and SplatSSC (51.83 mIoU, 62.83 IoU).
• On Occ-ScanNet-mini, AdaOcc with 500 queries achieves 57.97 mIoU and 65.16 IoU at 64.4 ms, versus 48.46 mIoU and 60.86 IoU at 85.1 ms for the best SplatSSC anchor setting.
• On Occ3D-nuScenes, AdaOcc with RGB+LiDAR reaches 50.89 RayIoU versus 41.2 for OPUS-L (8f).

Surprising findings

• Swapping the geometry prior inside GPOcc changes its IoU by 6.18 points, suggesting competitor gains are largely due to the prior rather than the architecture.
• Ground-truth depth improves AdaOcc only moderately (IoU 69.61 vs 65.50 with DAv2), indicating depth-estimator errors are a secondary bottleneck.

Limitations

• Reconstruction of small and thin objects in cluttered scenes is weak, as the authors note.
• Performance depends on the external monocular depth estimator; MoGe-2 lowers IoU by 3.06 points versus DAv2.
• Occ-ScanNet results rely on estimated depth rather than ground-truth geometry; LiDAR variant on TartanGround is lower (55.65 IoU) than GT-depth (67.55).

Experimental setup

Datasets: Occ-ScanNet Occ-ScanNet-mini TartanGround Occ3D-nuScenes

Benchmarks: Occ-ScanNet Occ-ScanNet-mini TartanGround Occ3D-nuScenes FloorPlan-R2R

Metrics: IoU mIoU per-class IoU RayIoU AbsRel RMSE NE OSR SR SPL

Baselines: TPVFormer MonoScene ISO SurroundOcc GaussianFormer OPUS AdaSFormer EmbodiedOcc EmbodiedOcc++ DiScene RoboOcc SplatSSC

Builds on: OPUS SplatSSC Depth Anything V2 DCD

Closest papers by meaning

• ViPOcc: Leveraging Visual Priors from Vision Foundation Models for Single-View 3D Occupancy Prediction (AAAI 2025 · similarity 0.90)
• OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.90)
• 3D Occupancy Prediction with Low-Resolution Queries via Prototype-aware View Transformation (CVPR 2025 · similarity 0.90)
• Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.90)
• SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction (CVPR 2024 · similarity 0.89)
• LowRankOcc: Tensor Decomposition and Low-Rank Recovery for Vision-based 3D Semantic Occupancy Prediction (CVPR 2024 · similarity 0.89)