Research map / Image generation / Editing and style transfer
Image-to-image translation for cross-domain disentanglement
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 |
|---|---|---|---|
| Use shared representation instead of pixels for retrieval | Cross-domain retrieval · MNIST-CD/CB | Recall@1: 30.45 -> 99.99; 40.02 -> 99.95 | Helped |
| Use shared features instead of BicycleGAN retrieval | Cross-domain retrieval · Facades and Maps | Recall@1: 45 -> 97; 68.0 -> 96.9 | Helped |
| Use cross-domain disentanglement rather than baseline translation | Many-to-many image translation · Cars and Chairs | LPIPS: Cars: 9.39$\pm$2.3 -> \textbf{8.30$\bm{\pm}$2.5}; 7.27$\pm$1.5 -> \textbf{4.66 $\bm{\pm}$ 1.1}; Chairs: 9.60$\pm$3.0 -> | Mixed |
| Remove cross-domain autoencoders (ablation) | Cross-domain disentanglement model · Visual analogies on MNIST, Cars, and Chairs | Distance: 13.0 $\pm$ 3.2 -> 35.1$\pm$13.2; 10.2 $\pm$ 2.6 -> 15.0 $\pm$ 4.6; 8.9 $\pm$ 2.8 -> 18.3 $\pm$ 2.4 | Hurt |
| Replace cross-domain autoencoders with normal autoencoders (ablation) | Cross-domain disentanglement model · Visual analogies on MNIST and Cars | Distance: 13.0 $\pm$ 3.2 -> 40.0 $\pm$ 11.5; 10.2 $\pm$ 2.6 -> 16.8 $\pm$ 5.04; 8.9 $\pm$ 2.8 -> 9.3 $\pm$ 1.4 | Hurt |
| Remove shared-representation L1 loss (ablation) | Cross-domain disentanglement model · Visual analogies on MNIST and Cars | Distance: 13.0 $\pm$ 3.2 -> 16.6 $\pm$ 5.9; 5.4 $\pm$ 1.6 -> 18.0 $\pm$ 2.5 | Hurt |
| Remove latent reconstruction loss (ablation) | Cross-domain disentanglement model · Visual analogies on MNIST-CB and Chairs-P | Distance: 13.0 $\pm$ 3.2 -> 18.9$\pm$6.4; 11.4 $\pm$ 2.4 -> 15.4 $\pm$ 1.5 | Hurt |
| Remove the Gradient Reversal Layer (ablation) | Exclusive representation · MNIST intra-domain retrieval using exclusive features | Recall increase indicating shared information: 0.7%; 8%; 25% | Mixed |
Closest papers by meaning
- A Unified Feature Disentangler for Multi-Domain Image Translation and Manipulation (NeurIPS 2018 · similarity 0.93)
- Multi-mapping Image-to-Image Translation via Learning Disentanglement (NeurIPS 2019 · similarity 0.93)
- Detach and Adapt: Learning Cross-Domain Disentangled Deep Representation (CVPR 2018 · similarity 0.89)
- Variational Interaction Information Maximization for Cross-domain Disentanglement (NeurIPS 2020 · similarity 0.89)
- One-Shot Unsupervised Cross Domain Translation (NeurIPS 2018 · similarity 0.88)
- Homomorphic Latent Space Interpolation for Unpaired Image-To-Image Translation (CVPR 2019 · 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.