Research map / 3D vision / Face and shape
Matryoshka Networks: Predicting 3D Geometry via Nested Shape Layers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace a dense voxel decoder with a voxel-tube image decoder | ShapeNet-core single-image reconstruction · ShapeNet-core 13 categories | mean IoU (%): 56.0 -> 64.1 | Helped |
| use nested shape layers instead of a dense voxel-tube output | ShapeNet-core single-image reconstruction · ShapeNet-core 13 categories | mean IoU (%): 64.1 -> 63.5 | Mixed |
| increase output resolution | Matryoshka network · ShapeNet-cars 256³ ground truth | IoU (%): 68.3 -> 78.4 -> 79.4 -> 79.6 | Helped |
| increase the number of nested shape layers (ablation) | ShapeNet-cars 128³ shape encoding · ShapeNet-cars 128³ | completely reconstructed (%): 2.6 -> 35.2 -> 94.3 -> 99.9 -> 100.0 | Helped |
| use a ResNet-inspired base architecture (ablation) | voxel tube network · ShapeNet-core car, chair, table | mean IoU (%): 59.2 -> 63.8 | Helped |
| remove residual connections (ablation) | voxel tube network · ShapeNet-core car, chair, table | mean IoU (%): 63.8 -> 60.8 | Hurt |
| use U-Net instead of the ResNet-inspired architecture (ablation) | voxel tube network · ShapeNet-core car, chair, table | mean IoU (%): 63.8 -> 62.6 | Hurt |
| train with approximate IoU loss instead of binary cross-entropy (ablation) | voxel tube network · ShapeNet-core car, chair, table | mean IoU (%): 64.0 -> 64.7 | Helped |
| train with cosine similarity loss instead of binary cross-entropy (ablation) | voxel tube network · ShapeNet-core car, chair, table | mean IoU (%): 64.0 -> 64.5 | Helped |
| train with L1 loss instead of binary cross-entropy (ablation) | voxel tube network · ShapeNet-core car, chair, table | mean IoU (%): 64.0 -> 62.7 | Hurt |
| use the Matryoshka network rather than OGN | single-image reconstruction · ShapeNet-cars 256³ ground truth | IoU (%): 76.6 -> 79.6 | Helped |
| use Matryoshka shape-from-ID rather than OGN | Blendswap shape reconstruction from ID · Blendswap 512³ | IoU (%): 96.9% -> 97.8% | Helped |
Closest papers by meaning
- Single-View 3D Object Reconstruction From Shape Priors in Memory (CVPR 2021 · similarity 0.88)
- RIM-Net: Recursive Implicit Fields for Unsupervised Learning of Hierarchical Shape Structures (CVPR 2022 · similarity 0.87)
- Geometry-Aware Network for Non-Rigid Shape Prediction From a Single View (CVPR 2018 · similarity 0.87)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.87)
- Training-free Spatially Grounded Geometric Shape Encoding (NeurIPS 2026 · similarity 0.87)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · 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.