Research map / Detection and segmentation / Open-set and few-shot detection
Unsupervised Robust Disentangling of Latent Characteristics for Image Synthesis
Patrick Esser, Johannes Haux, Bjorn Ommer
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 adaptive KL weight controlled by second classifier T' to adversarial disentanglement (ablation) | adversarial disentangling · Table 1 MI constraint 0.125 | I_T' (entanglement): 0.8727 -> 0.1201 | Improved |
| add adaptive KL weight instead of fixed gamma=1 (ablation) | our method · Table 1 MI constraint 0.125 | I_T' (entanglement): 0.4096 -> 0.1201 | Improved |
| add adaptive weighting instead of only variational regularization (ablation) | our method · Table 1 MI constraint 0.125 | reconstruction loss: 3.9350 -> 3.5279 | Improved |
| add KL term to adversarial approach (reconstruction cost) (ablation) | adversarial only · Table 1 MI constraint 0.125 | reconstruction loss: 2.9818 -> 3.5279 | Hurt |
| combine KL and MI with equal weights rather than adaptive (ablation) | our method · Table 1 MI constraint 0.125 | reconstruction loss: 3.9615 -> 3.5279 | Improved |
| add KL term (L_vb) to adversarial model on Market-1501 transfer (ablation) | adversarial · Market-1501 (transfer) | reID mAP (%): 6.6 -> 25.4 | Improved |
| add KL term (L_vb) to adversarial model on DeepFashion transfer (ablation) | adversarial · DeepFashion (transfer) | reID mAP (%): 10.5 -> 47.1 | Improved |
| add KL term to adversarial model, Market-1501 reconstruction (ablation) | adversarial · Market-1501 (reconstruction) | reID mAP (%): 34.8 -> 30.2 | Hurt |
| add KL term to adversarial model, DeepFashion pose error (ablation) | adversarial · DeepFashion (transfer) | pose error (% of image width): 7.4 +- 8.1 -> 6.8 +- 7.6 | Improved |
| use unsupervised method instead of supervised VUNet, DeepFashion transfer appearance | VUNet · DeepFashion (transfer) | reID mAP (%): 14.6 -> 47.1 | Improved |
| use unsupervised method instead of keypoint-supervised VUNet, pose error on Market-1501 transfer | VUNet · Market-1501 (transfer) | pose error (% of image width): 17.5 +- 9.4 -> 24.7 +- 9.9 | Hurt |
Closest papers by meaning
- Explicit Disentanglement of Appearance and Perspective in Generative Models (NeurIPS 2019 · similarity 0.91)
- Unsupervised Part-Based Disentangling of Object Shape and Appearance (CVPR 2019 · similarity 0.90)
- Overcoming the Disentanglement vs Reconstruction Trade-off via Jacobian Supervision (ICLR 2019 · similarity 0.90)
- An Adversarial Neuro-Tensorial Approach for Learning Disentangled Representations (IJCV 2019 · similarity 0.90)
- Scaling-Up Disentanglement for Image Translation (ICCV 2021 · similarity 0.90)
- Disentangled3D: Learning a 3D Generative Model With Disentangled Geometry and Appearance From Monocular Images (CVPR 2022 · similarity 0.89)
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.
