Research map / 3D vision / Face and shape
Neural 3D Morphable Models: Spiral Convolutional Networks for 3D Shape Representation Learning and Generation
Giorgos Bouritsas, Sergiy Bokhnyak, Stylianos Ploumpis, Michael Bronstein, Stefanos Zafeiriou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use fixed consistent spiral ordering instead of random per mesh and epoch (linear projection) (ablation) | Spiral convolution · COMA (ordering consistency, simple Neural3DMM, latent size 16) | reconstruction error (mm): 0.829 -> 0.635 | Improved |
| use fixed consistent spiral ordering instead of random per mesh and epoch (LSTM) (ablation) | Spiral LSTM operator · COMA (ordering consistency, simple Neural3DMM, latent size 16) | reconstruction error (mm): 0.888 -> 0.792 | Improved |
| use fixed ordering instead of random orientation per epoch (linear projection) (ablation) | Spiral convolution · COMA (ordering consistency, simple Neural3DMM, latent size 16) | reconstruction error (mm): 0.951 -> 0.635 | Improved |
| replace LSTM over the spiral with linear projection (fixed ordering) (ablation) | Spiral operator · COMA (ordering consistency, simple Neural3DMM, latent size 16) | reconstruction error (mm): 0.792 -> 0.635 | Improved |
| replace MoNet 9 kernels with spiral convolution (ablation) | MoNet 9 kernels · COMA (latent size 16, simple Neural3DMM architecture) | reconstruction error (mm): 0,708 -> 0,635 | Improved |
| replace GAT 9 kernels with spiral convolution (ablation) | GAT 9 kernels · COMA (latent size 16, simple Neural3DMM architecture) | reconstruction error (mm): 0,762 -> 0,635 | Improved |
| replace FeastNet 9 kernels with spiral convolution (ablation) | FeastNet 9 kernels · COMA (latent size 16, simple Neural3DMM architecture) | reconstruction error (mm): 0,750 -> 0,635 | Improved |
| replace over-parametrised MoNet 25 kernels with spiral convolution (ablation) | MoNet 25 kernels · COMA (latent size 16, simple Neural3DMM architecture) | reconstruction error (mm): 0,583 -> 0,635 | Hurt |
| replace over-parametrised FeastNet 25 kernels with spiral convolution (ablation) | FeastNet 25 kernels · COMA (latent size 16, simple Neural3DMM architecture) | reconstruction error (mm): 0,623 -> 0,635 | Hurt |
| replace over-parametrised GAT 25 kernels with spiral convolution (ablation) | GAT 25 kernels · COMA (latent size 16, simple Neural3DMM architecture) | reconstruction error (mm): 0,732 -> 0,635 | Improved |
| replace MoNet 9 kernels with spiral convolution (ablation) | MoNet 9 kernels · COMA (latent size 16, simple Neural3DMM architecture) | inference time per mesh (ms): 10,55 -> 8,18 | Improved |
| replace spectral ChebNet filters with spiral convolution (ablation) | COMA autoencoder · COMA and DFAUST (varying latent sizes) | generalisation error | Improved |
Closest papers by meaning
- Adaptive Spiral Layers for Efficient 3D Representation Learning on Meshes (ICCV 2023 · similarity 0.91)
- FeaStNet: Feature-Steered Graph Convolutions for 3D Shape Analysis (CVPR 2018 · similarity 0.90)
- Learning Feature Aggregation for Deep 3D Morphable Models (CVPR 2021 · similarity 0.89)
- Pixel2Mesh: 3D Mesh Model Generation via Image Guided Deformation (TPAMI 2020 · similarity 0.89)
- Learning Local Neighboring Structure for Robust 3D Shape Representation (AAAI 2021 · similarity 0.88)
- Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space (CVPR 2025 · similarity 0.88)
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.
