Research map / 3D vision / Face and shape
$\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation
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 NAISR, a neural additive implicit shape model that deforms a SIREN-based SDF atlas by a sum of covariate-specific displacement fields plus a latent code, enabling reconstruction, disentanglement, evolution and transfer of 3D shapes.
Problem
Existing deep implicit shape representations cannot precisely represent shape populations while disentangling and capturing individual covariate effects (e.g., age, weight, sex, disease) on shape, which is needed for scientific shape discovery.
Approach
Represent each shape as a template SDF warped by a sum of covariate-controlled displacement fields (each zeroed at the mean covariate value) plus a shared latent code, making the deformation additive over covariates like a neural additive model, so individual covariate effects can be isolated, extrapolated and transferred.
Key results
- On the pediatric airway dataset, NAISR (inferred covariates) achieves mean Chamfer distance 0.067 vs 0.094 for the best baseline DIT (x10^2 units in the table), and with covariates given 0.084.
- On shape transfer, NAISR reduces mean volume difference on ADNI hippocampus to 12.82 cm^3 (no covariates) and 11.23 cm^3 (with covariates) vs 81.07 for A-SDF without covariates on hippocampus; on airway, 8.84 vs 40.96-82.92 for A-SDF.
- On Starman reconstruction, A-SDF with covariates performs best (CD 0.049) with NAISR with covariates matching it (CD 0.049).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Supply covariates during inference instead of inferring them | NAISR · Starman test | Chamfer distance: 0.111 -> 0.049 | Helped |
| Supply covariates during inference instead of inferring them | NAISR · Starman test | Earth mover's distance: 1.709 -> 1.276 | Helped |
| Supply covariates during inference instead of inferring them | NAISR · ADNI Hippocampus test | Chamfer distance: 0.174 -> 0.126 | Helped |
| Supply covariates during inference instead of inferring them | NAISR · Pediatric Airway test | Chamfer distance: 0.067 -> 0.084 | Hurt |
| Use NAISR for shape transfer rather than A-SDF with covariates | Shape transfer · ADNI Hippocampus | Volume difference: 0.215 -> 0.086 | Helped |
| Use NAISR for shape transfer rather than A-SDF with covariates | Shape transfer · Pediatric Airway | Volume difference: 41.46 -> 11.23 | Helped |
| Use NAISR rather than the strongest listed baseline | Shape reconstruction · Starman test | Chamfer distance: A-SDF (c) 0.049; NAISR (c) 0.049 | Mixed |
| Use NAISR rather than the strongest listed baseline | Shape reconstruction · ADNI Hippocampus test | Chamfer distance: DIT 0.156; NAISR (c) 0.126 | Helped |
| Use NAISR rather than the strongest listed baseline | Shape reconstruction · Pediatric Airway test | Earth mover's distance: DeepSDF 1.401; NAISR 1.233 | Helped |
Surprising findings
- A-SDF, which conditions on covariates, performs poorly on real 3D medical shapes, plausibly memorizing training shapes via covariates rather than disentangling effects.
- The NAISR template-based deformation is not guaranteed to be invertible, which the authors flag as a limitation.
Limitations
- Deformations are not guaranteed invertible, so point correspondences may not preserve topology.
- Only indirect assessment of the learned model via reconstruction and transfer; no direct assessment of abnormality detection.
- Real datasets are private or limited to normal airways; the model is applied only to normal airway anatomy.
Experimental setup
Datasets: Starman (synthetic 2D) ADNI hippocampus Pediatric airway CT segmentations
Benchmarks: Starman ADNI hippocampus Pediatric airway
Metrics: Chamfer distance earth mover's distance Hausdorff distance volume difference
Baselines: DeepSDF A-SDF DIT NDF
Builds on: DeepSDF SIREN neural additive models A-SDF
Closest papers by meaning
- Shape Unicode: A Unified Shape Representation (CVPR 2019 · similarity 0.88)
- Implicit Modeling of Non-rigid Objects with Cross-Category Signals (AAAI 2024 · similarity 0.87)
- Do 2D GANs Know 3D Shape? Unsupervised 3D Shape Reconstruction from 2D Image GANs (ICLR 2021 · similarity 0.86)
- TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding (CVPR 2024 · similarity 0.86)
- 3D Shape Reconstruction From 2D Images With Disentangled Attribute Flow (CVPR 2022 · similarity 0.86)
- ShapeEmbed: a self-supervised learning framework for 2D contour quantification (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.