Research map / Point clouds / Shape and surface completion
DITTO: Dual and Integrated Latent Topologies for Implicit 3D Reconstruction
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 DITTO, which jointly refines point and grid latents with a dual latent layer (DSPT point transformer) and decodes them with an integrated implicit decoder for implicit 3D reconstruction from noisy, sparse point clouds.
Problem
Implicit 3D reconstruction from noisy and sparse point clouds: grid latents are stable but lack detail due to resolution limits, while point latents preserve detail but are unstable to noise and produce holes in thin structures.
Approach
Keep point and grid latents as separate parallel representations refined recursively within a dual latent layer (with a dynamic sparse point transformer on points), then decode with an integrated decoder that combines KNN point neighbors with grid-interpolated features via self-attention, instead of alternating projections or a grid-only decoder as in ALTO.
Key results
- On ShapeNet with 3K points, DITTO reaches IoU 0.949 vs 0.930 for ALTO (previous SoTA).
- On Synthetic Rooms (10K points, 0.005 noise), DITTO reaches IoU 0.928 and Chamfer-L1 0.34 vs ALTO's 0.914 and 0.35.
- With triplane grid latents on Synthetic Rooms, DITTO attains IoU 0.931 while ALTO drops to 0.895.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DITTO instead of ALTO | ShapeNet, normal (3K points) · ShapeNet, normal (3K points) | IoU: 0.930 -> 0.949 | Helped |
| Use DITTO instead of ALTO | ShapeNet, sparse (1K points) · ShapeNet, sparse (1K points) | IoU: 0.905 -> 0.926 | Helped |
| Use DITTO instead of ALTO | ShapeNet, sparse (300 points) · ShapeNet, sparse (300 points) | IoU: 0.863 -> 0.882 | Helped |
| Use DITTO instead of ALTO | Synthetic Rooms, 10K points · Synthetic Rooms, 10K points | IoU: 0.914 -> 0.928 | Helped |
| Add DLL with DSPT backbone to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.921 | Helped |
| Add DLL with FKAConv backbone to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.917 | Helped |
| Add DLL with PointTransformer backbone to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.911 | Helped |
| Add IID alone to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.907 | Helped |
| Add FKAConv PointEncoder to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.912 | Helped |
| Add DLL and IID to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.925 | Helped |
| Add DLL and PointEncoder to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.926 | Helped |
| Add IID and PointEncoder to ALTO (ablation) | Synthetic Rooms triplane baseline · Synthetic Rooms, 10K points | IoU: 0.895 -> 0.918 | Helped |
Surprising findings
- Adding IID alone slightly lowers NC and F-score relative to ALTO, and gains only materialize when combined with DLL.
Limitations
- Triplane grid latents generalize poorly to out-of-distribution data such as ScanNet-v2, as triplanes lack geometric inductive bias.
- Sparse-point and noisy-input gains over ALTO are modest in some settings (e.g., noisy input IoU 0.811 vs 0.804, same Chamfer and F-score).
Experimental setup
Datasets: ShapeNet Synthetic Rooms ScanNet-v2
Benchmarks: ShapeNet Synthetic Rooms ScanNet-v2
Metrics: IoU Chamfer-L1 normal consistency F-score
Baselines: ONet ConvONet POCO ALTO DP-ConvONet SPSR
Builds on: ALTO ConvONet POCO FKAConv DSVT
Closest papers by meaning
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.87)
- Triplane Meets Gaussian Splatting: Fast and Generalizable Single-View 3D Reconstruction with Transformers (CVPR 2024 · similarity 0.86)
- DOCTR: Disentangled Object-Centric Transformer for Point Scene Understanding (AAAI 2024 · similarity 0.85)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.85)
- DIFIX3D+: Improving 3D Reconstructions with Single-Step Diffusion Models (CVPR 2025 · similarity 0.85)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 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.