Research map / 3D vision / Face and shape
Octree Generating Networks: Efficient Convolutional Architectures for High-Resolution 3D Outputs
Maxim Tatarchenko, Alexey Dosovitskiy, Thomas Brox
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 known octree structure at test time (Prop-known) instead of predicted (ablation) | OGN autoencoder · ShapeNet-cars autoencoder 64^3 | IoU: 0.884 -> 0.904 | Improved |
| fine-tune with Prop-pred after Prop-known training (ablation) | OGN autoencoder 2^3 filters · ShapeNet-cars autoencoder 64^3 | IoU: 0.862 -> 0.884 | Improved |
| fine-tune with Prop-pred after Prop-known training (ablation) | OGN autoencoder 4^3 filters · ShapeNet-cars autoencoder 64^3 | IoU: 0.804 -> 0.885 | Improved |
| use 4x4 up-convolutions instead of 2x2 (ablation) | OGN autoencoder · ShapeNet-cars autoencoder 64^3 | IoU: 0.884 -> 0.885 | No clear effect |
| interleave 3x3 convolutions (IntConv) (ablation) | OGN autoencoder · ShapeNet-cars autoencoder 64^3 | IoU: 0.884 -> 0.885 | No clear effect |
| replace dense decoder with OGN (memory) | decoder at 512^3 · Decoder efficiency | memory (GB): 74.28 -> 0.88 | Improved |
| replace dense decoder with OGN (time) | decoder at 256^3 · Decoder efficiency | iteration time (s): 3.89 -> 0.64 | Improved |
| replace dense decoder with OGN decoder | single-view reconstruction · ShapeNet-all 32^3 | mean IoU: 0.590 -> 0.596 | No clear effect |
| increase output resolution 64^3 to 128^3 (ablation) | OGN shape from ID · ShapeNet-cars | IoU: 0.856 -> 0.901 | Improved |
| increase output resolution 128^3 to 256^3 (ablation) | OGN shape from ID · ShapeNet-cars (full) | IoU: 0.901 -> 0.865 | Hurt |
| increase output resolution 128^3 to 256^3 on reduced dataset (ablation) | OGN shape from ID · ShapeNet-cars (500 models) | IoU: 0.922 -> 0.931 | Improved |
| increase output resolution 64^3 to 512^3 (ablation) | OGN shape from ID · BlendSwap | IoU: 0.535 -> 0.969 | Improved |
Closest papers by meaning
- OctNet: Learning Deep 3D Representations at High Resolutions (CVPR 2017 · similarity 0.90)
- Efficient Autoregressive Shape Generation via Octree-Based Adaptive Tokenization (ICCV 2025 · similarity 0.87)
- OctField: Hierarchical Implicit Functions for 3D Modeling (NeurIPS 2021 · similarity 0.87)
- Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.86)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.85)
- OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · similarity 0.84)
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.
