Research map / 3D vision / Face and shape
Learning to Generate Chairs, Tables and Cars with Convolutional Networks
Alexey Dosovitskiy, Jost Tobias Springenberg, Maxim Tatarchenko, 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 |
|---|---|---|---|
| add convolutions between up-convolutions (deeper network) (ablation) | 1s-S generator · Chairs | MSE (x10^-3): 3.51 -> 2.90 | Improved |
| widen up-convolutional layers by 1.3x (ablation) | 1s-S generator · Chairs | MSE (x10^-3): 3.51 -> 3.41 | No clear effect |
| use two separate streams for RGB and segmentation (ablation) | 1s-S generator · Chairs | MSE (x10^-3): 3.51 -> 3.44 | No clear effect |
| use squared Euclidean mask loss instead of NLL (2 streams) (ablation) | 2s-S · Chairs | MSE (x10^-3): 3.44 -> 3.43 | No clear effect |
| increase number of car models from 500 to 7124 (ablation) | 1s-S-deep · Cars | MSE (x10^-3): 0.48 -> 0.97 | Hurt |
| add data augmentation at 1000 car models (ablation) | 1s-S-deep · Cars | MSE (x10^-3): 0.66 -> 1.18 | Hurt |
| add data augmentation to improve smooth interpolation between styles (ablation) | 1s-S-deep · Cars (style morphing) | interpolation smoothness | Improved |
| train with all viewpoints of source chairs (knowledge transfer) (ablation) | 1s-S-deep 64x64 · Chairs (unseen viewpoints) | squared error of generated views | Improved |
| train on chairs and tables together (ablation) | generator · Chairs (unseen elevations) | view generation quality | Improved |
| use generated morphing sequences with optical flow instead of direct matching (all pairs) | 1s-S-deep · Chair correspondence | displacement (px): 4.0 -> 3.4 | Improved |
| use generated morphing instead of SIFT flow on difficult pairs | 1s-S-deep · Chair correspondence (difficult) | displacement (px): 4.8 -> 3.5 | Improved |
| use generated morphing instead of SIFT flow on simple pairs | 1s-S-deep · Chair correspondence (simple) | displacement (px): 2.8 -> 3.1 | Hurt |
Closest papers by meaning
- Multi-view 3D Models from Single Images with a Convolutional Network (ECCV 2016 · similarity 0.88)
- Visual Object Networks: Image Generation with Disentangled 3D Representations (NeurIPS 2018 · similarity 0.87)
- Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling (NeurIPS 2016 · similarity 0.87)
- Learning FRAME Models Using CNN Filters (AAAI 2016 · similarity 0.87)
- Warp and Learn: Novel Views Generation for Vehicles and Other Objects (TPAMI 2020 · similarity 0.86)
- Leveraging 2D Data to Learn Textured 3D Mesh Generation (CVPR 2020 · 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.
