Research map / Point clouds / Shape and surface completion
Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries
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.
Adds locally-radiused adversarial query points to NeuralPull-style unsupervised SDF learning from sparse unoriented point clouds, improving SDF reconstruction on ShapeNet, Faust, 3D Scene and SRB.
Problem
Learning neural signed distance functions from sparse, noisy, unoriented 3D point clouds without ground-truth supervision overfits to pseudo-label noise, causing shape degradation and unstable validation.
Approach
Augment NeuralPull's query-projection loss with a loss on local adversarial queries (first-order gradient-ascent perturbations within radii tied to local point spacing sigma_p), combined via learnable multi-task uncertainty weights, to regularize training and focus on hard samples.
Key results
- On ShapeNet sparse noisy inputs, the method reaches CD1 0.76 and CD2 0.020 versus 1.16 and 0.074 for NeuralPull.
- On Faust, the method reports CD1 0.220 and FS 0.981, outperforming NeuralPull (CD1 0.269, FS 0.973) and the supervised NKSR/SAC comparators in CD1.
- On the 3D Scene benchmark, mean CD1 improves from 0.097 (NP) to 0.041 for the proposed method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local adversarial query loss to the NeuralPull projection loss (ablation) | NP · ShapeNet sparse noisy | CD1 | Helped |
| add local adversarial query loss to the NeuralPull projection loss (ablation) | NP · Faust sparse noisy | CD1: 0.269 -> 0.220 | Helped |
| add local adversarial query loss to the NeuralPull projection loss (ablation) | NP · 3D Scene mean | CD1: 0.097 -> 0.041 | Helped |
| add local adversarial query loss to the NeuralPull projection loss (ablation) | NP · SRB sparse | CD1: 0.58 -> 0.49 | Helped |
| add local adversarial query loss to the NeuralPull projection loss (ablation) | NP · SRB dense | CD1: 0.23 -> 0.19 | Helped |
| use local radius ρq = σp/100 rather than global radius ρ = σ/100 (ablation) | Ours · ShapeNet Table class | CD1: 0.98 -> 0.75 | Helped |
| set the local adversarial radius to σp/10 rather than σp/100 (ablation) | Ours · ShapeNet Table class | CD1: 0.75 -> 0.92 | Hurt |
| set the local adversarial radius to σp/500 rather than σp/100 (ablation) | Ours · ShapeNet Table class | CD1: 0.75 -> 0.88 | Hurt |
| set the local adversarial radius to σp/1000 rather than σp/100 (ablation) | Ours · ShapeNet Table class | CD1: 0.75 -> 1.02 | Hurt |
| replace the combined original and adversarial objective with adversarial loss alone (ablation) | Ours · ShapeNet Table class | CD1: 0.75 -> 0.78 | Hurt |
| increase NP's original query samples to match the total number of samples used by Ours (ablation) | NP · SRB | Chamfer distance: 0.581 to 0.576 | Helped |
Surprising findings
- Increasing NeuralPull's number of query points to match the total of the proposed method yields only marginal gains (SRB average CD 0.581 to 0.576).
Limitations
- Struggles to learn very thin and convoluted structures such as chair legs and slats.
- Balancing empirical risk minimization and adversarial loss relies on an off-the-shelf weighting strategy.
- Very large adversarial radii can produce spurious pseudo supervision by changing nearest input point.
Experimental setup
Datasets: ShapeNet Faust 3D Scene Surface Reconstruction Benchmark (SRB)
Benchmarks: ShapeNet Faust 3D Scene SRB
Metrics: CD1 CD2 F-Score Normal Consistency
Baselines: NeuralPull SPSR OG-INR GridPull NTPS NDrop SAP DIGS NSpline POCO CONet NKSR
Builds on: NeuralPull adversarial training (local adversarial perturbations) Multi-task learning uncertainty weighting (Kendall et al.)
Closest papers by meaning
- Unsupervised Occupancy Learning from Sparse Point Cloud (CVPR 2024 · similarity 0.89)
- Small Steps and Level Sets: Fitting Neural Surface Models with Point Guidance (CVPR 2024 · similarity 0.88)
- Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes (CVPR 2024 · similarity 0.88)
- Sharpening Neural Implicit Functions with Frequency Consolidation Priors (AAAI 2025 · similarity 0.88)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.87)
- Toward Robust Neural Reconstruction from Sparse Point Sets (CVPR 2025 · similarity 0.87)
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.