Research map / 3D vision / Face and shape

DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models

NeurIPS 2026 · 3D vision › Face and shape · approach: Single-view 3D reconstruction · 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.

Distills DINOv3 features into a TSDF-only student network and fuses them with TSDF features via a multi-scale voxel Mamba for deterministic 3D shape completion on unseen categories.

Problem

Shape completion from partial TSDF scans struggles on unseen categories and noisy real-world scans, where geometry alone is insufficient to infer missing structure, and existing strong methods are either costly (diffusion) or heavy (patch-based).

Approach

Train a TSDF-only student-teacher model that predicts voxel-aligned DINOv3 feature volumes from partial geometry, and fuse these semantic features with TSDF features using a voxel state space operator plus a multi-scale (full-grid and chunk-wise) Mamba refinement.

Key results

• On unseen ShapeNet categories, average CD/IoU of 3.64 / 71.1 versus DiffComplete 4.10 / 67.5 and PatchComplete 4.27 / 65.4.
• On ScanNet unseen categories, average CD/IoU of 7.09 / 53.0 versus DiffComplete 7.18 / 51.3.
• Model has 25.97M params, 0.292 GB memory, and 28.68 ms inference time per sample, versus DiffComplete 43.09M, 3115.85 ms and PatchComplete 182.85M, 65.56 ms.

Surprising findings

• Directly injecting ground-truth DINOv3 features at inference (3.27 CD / 75.4 IoU) outperforms the distilled TSDF-only features, indicating distillation quality is a bottleneck.
• DiffComplete degrades sharply with reduced training data (19.46 CD / 11.8 IoU at 20%), while the proposed method stays robust.

Limitations

• Gap between distilled TSDF-DINO features and ground-truth DINO features limits performance.
• Trained mainly on synthetic data; domain gap to noisy ScanNet scans degrades semantic feature quality.
• ScanNet results are only modestly better than baselines and the method is evaluated on a limited set of categories.

Experimental setup

Datasets: ShapeNet ScanNet Scan2CAD

Benchmarks: ShapeNet unseen categories (PatchComplete split) ScanNet objects (unseen categories) 3D-EPN known categories

Metrics: Chamfer Distance IoU l1 error parameters GPU memory inference time

Baselines: 3D-EPN Few-Shot completion IF-Nets AutoSDF PatchComplete DiffComplete SDF-StyleGAN RePaint-3D

Builds on: DINOv3 VoxelMamba PatchComplete Mamba DiffComplete

Closest papers by meaning

• PatchComplete: Learning Multi-Resolution Patch Priors for 3D Shape Completion on Unseen Categories (NeurIPS 2022 · similarity 0.89)
• Voxel Proposal Network via Multi-Frame Knowledge Distillation for Semantic Scene Completion (NeurIPS 2024 · similarity 0.87)
• Memory-Augmented Re-Completion for 3D Semantic Scene Completion (AAAI 2025 · similarity 0.86)
• Context and Geometry Aware Voxel Transformer for Semantic Scene Completion (NeurIPS 2024 · similarity 0.86)
• Learning Generalizable Shape Completion with SIM(3) Equivariance (NeurIPS 2025 · similarity 0.86)
• Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation (CVPR 2024 · similarity 0.86)