Research map / Point clouds / Shape and surface completion
Sharpening Neural Implicit Functions with Frequency Consolidation Priors
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.
Learns frequency consolidation priors that map low-frequency SDF observations to full-frequency SDFs via disentangled embeddings, enabling sharpening of neural implicit surfaces at test time via self-reconstruction.
Problem
Neural SDFs learned from sparse point clouds or multi-view images struggle to represent high-frequency geometry such as sharp edges and structures.
Approach
Learn a data-driven mapping from low-frequency SDF observations to full-frequency SDFs, representing frequency components as embeddings and disentangling the full-frequency shape embedding from the low-frequency corruption embedding; at test time only the embedding is optimized via self-reconstruction on the fixed low-frequency decoder.
Key results
- On ShapeNet (13 classes), the method reaches CD_L1x10 0.187 and CD_L2x100 0.216 versus 0.213 and 0.285 for the next best ALTO.
- Applied to sparse point cloud reconstructions from NeuralTPS, it lowers CD_L1x10 from 0.141 to 0.115 and raises NC from 0.899-era baseline to 0.918.
- On ABC CAD shapes, mean CD_L1 is 0.011 versus 0.014 for Points2Surf.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply frequency consolidation priors to the ShapeNet reconstruction method | ShapeNet 13-class reconstruction · ShapeNet 13 classes | CD_L1×10: 0.213 -> 0.187 | Helped |
| Apply frequency consolidation priors to the ShapeNet reconstruction method | ShapeNet 13-class reconstruction · ShapeNet 13 classes | CD_L2×100: 0.284 -> 0.216 | Helped |
| Apply frequency consolidation priors to the ShapeNet reconstruction method | ShapeNet 13-class reconstruction · ShapeNet 13 classes | NC: 0.861 -> 0.871 | Helped |
| Apply frequency consolidation priors to OnSurf reconstructions | OnSurf · ShapeNet sparse point cloud reconstruction (300 points, 8 classes) | CD_L1×10: 0.214 -> 0.180 | Helped |
| Apply frequency consolidation priors to OnSurf reconstructions | OnSurf · ShapeNet sparse point cloud reconstruction (300 points, 8 classes) | CD_L2×100: 0.223 -> 0.165 | Helped |
| Apply frequency consolidation priors to OnSurf reconstructions | OnSurf · ShapeNet sparse point cloud reconstruction (300 points, 8 classes) | NC: 0.845 -> 0.886 | Helped |
| Apply frequency consolidation priors to NeuralTPS reconstructions | NeuralTPS · ShapeNet sparse point cloud reconstruction (300 points, 8 classes) | CD_L1×10: 0.141 -> 0.115 | Helped |
| Apply frequency consolidation priors to NeuralTPS reconstructions | NeuralTPS · ShapeNet sparse point cloud reconstruction (300 points, 8 classes) | CD_L2×100: 0.093 -> 0.088 | Helped |
| Apply frequency consolidation priors to NeuralTPS reconstructions | NeuralTPS · ShapeNet sparse point cloud reconstruction (300 points, 8 classes) | NC: 50.899 -> 0.918 | Helped |
| Train and use frequency consolidation priors for reconstruction | ABC reconstruction · ABC | CD_L1 mean: 0.014 -> 0.011 | Helped |
| Train and use frequency consolidation priors for reconstruction | ABC reconstruction · ABC | NC mean: 0.902 -> 0.962 | Helped |
Surprising findings
- Pretrained IDF and ConvOcc, when applied to corrupted shapes, can generate some high-frequency geometry but fail on large geometry variations.
Limitations
- Test-time optimization of embeddings per shape is required (800 iterations).
- Relies on surface reconstruction (NeuralPull) to convert point clouds or meshes into SDFs before applying the prior.
Experimental setup
Datasets: ShapeNet ABC ScanNet Scan2CAD
Benchmarks: ShapeNet ABC ScanNet
Metrics: CD_L1 CD_L2 normal consistency
Baselines: DeepSDF ConvOcc LIG IDF NDC POCO ALTO OnSurf NeuralTPS Points2Surf SECAD-Net NeuralPartPriors
Builds on: DeepSDF Poisson surface reconstruction NeuralPull auto-decoding
Closest papers by meaning
- Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries (ICML 2024 · similarity 0.88)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.87)
- Small Steps and Level Sets: Fitting Neural Surface Models with Point Guidance (CVPR 2024 · similarity 0.87)
- MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step (NeurIPS 2024 · similarity 0.86)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.86)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.86)
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.