Research map / Detection and segmentation / 3D detection and driving
LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset
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.
Introduces LASA, a dataset of 10,412 artist-made CAD models aligned with 920 ArkitScenes real scans, plus a diffusion-based cross-modal reconstruction method (DisCo) and occupancy-guided 3D detection (OccGOD) that use it.
Problem
Instance-level shape reconstruction from real indoor scans needs large, well-aligned paired CAD annotations, but existing real datasets (e.g., Scan2CAD) use misaligned retrieved CAD models and synthetic datasets have domain gaps.
Approach
Artist-crafted, geometrically and pixel-level verified CAD annotations on laser-scanned ArkitScenes scenes, plus a triplane latent diffusion model with a Hybrid Feature Aggregation layer that fuses partial point clouds and multi-view images via a volumetric grid, and an occupancy head added to CAGroup3D using scene-level occupancy derived from LASA.
Key results
- DisCo with point clouds plus images reaches 38.6 mIoU on chair, 70.7 on sofa and 75.1 on cabinet, outperforming IFNet, 3DShape2VecSet and LAS Diffusion on LASA.
- OccGOD improves CAGroup3D overall mAP@0.5 from 55.72 to 57.36 and mAR from 66.43 to 67.47 on LASA test set.
- Pretraining on synthetic data then fine-tuning on LASA gives the best real-world reconstruction (e.g., chair 38.6 mIoU vs 25.0 without fine-tuning).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-view images to point-cloud input (ablation) | DisCo · LASA test, Chair | Chamfer L2: 5.68 -> 3.57 | Helped |
| Add multi-view images to point-cloud input (ablation) | DisCo · LASA test, Sofa | Chamfer L2: 3.21 -> 2.88 | Helped |
| Fine-tune on LASA after synthetic pretraining (ablation) | DisCo · LASA test, Chair | mIoU: 25.0 -> 38.6 | Helped |
| Use synthetic pretraining before LASA fine-tuning (ablation) | DisCo · LASA test, Chair | mIoU: 35.6 -> 38.6 | Helped |
| Use the Hybrid Feature Aggregation layer instead of direct triplane projection (ablation) | DisCo · LASA test, Chair | Chamfer L2: 3.95 -> 3.57 | Helped |
| Use noisy rather than ground-truth 3D bounding boxes (ablation) | DisCo · LASA test, Chair | Chamfer L2: 3.57 -> 3.77 | Hurt |
| Add scene-level occupancy prediction | CAGroup3D · LASA test, overall | mAP: 55.72 -> 57.36 | Helped |
| Add scene-level occupancy prediction | CAGroup3D · LASA test, overall | mAR: 66.43 -> 67.47 | Helped |
Surprising findings
- Chamfer distance of Scan2CAD CAD annotations to scans is 0.269 versus 0.161 for LASA, showing retrieved CADs are less aligned.
- OccGOD slightly decreases AP on some classes such as TV Monitor (1.16 to 0.73) and Washer (87.77 to 86.27).
Limitations
- Reconstruction evaluated on only 6 merged categories out of 17 in LASA.
- Occupancy-guided detection gains are modest (+1.64 mAP overall).
Experimental setup
Datasets: LASA ShapeNet ABO 3D-Future HM3D ArkitScenes Scan2CAD
Benchmarks: LASA test set
Metrics: mIoU Chamfer L2 F-score mAP mAR
Baselines: IFNet 3DShape2VecSet LAS Diffusion CAGroup3D
Builds on: EDM 3DShape2VecSet CAGroup3D LAS Diffusion Triplane VAE
Closest papers by meaning
- 3D-RCNN: Instance-Level 3D Object Reconstruction via Render-and-Compare (CVPR 2018 · similarity 0.87)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.86)
- 3DCoMPaT200: Language Grounded Large-Scale 3D Vision Dataset for Compositional Recognition (NeurIPS 2024 · similarity 0.86)
- ImOV3D: Learning Open Vocabulary Point Clouds 3D Object Detection from Only 2D Images (NeurIPS 2024 · similarity 0.85)
- MESC-3D:Mining Effective Semantic Cues for 3D Reconstruction from a Single Image (CVPR 2025 · similarity 0.85)
- Hierarchical Cross-Modal Alignment for Open-Vocabulary 3D Object Detection (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.