Research map / Image generation / GANs
Finding an Unsupervised Image Segmenter in each of your Deep Generative Models
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 |
|---|---|---|---|
| Automatically discover a generator-specific foreground/background direction and train a segmenter on generated image-mask pairs | Pretrained GANs · DUTS | IoU | Helped |
| Combine foreground-light and foreground-dark latent directions rather than use either direction alone (ablation) | BigBiGAN segmentation pipeline · CUB | IoU: 0.601 -> 0.664 (vl); 0.631 -> 0.664 (vd) | Helped |
| Combine foreground-light and foreground-dark latent directions rather than use either direction alone (ablation) | BigBiGAN segmentation pipeline · DUTS | IoU: 0.486 -> 0.528 (vl); 0.478 -> 0.528 (vd) | Helped |
| Combine foreground-light and foreground-dark latent directions rather than use either direction alone (ablation) | BigBiGAN segmentation pipeline · ECSSD | IoU: 0.663 -> 0.713 (vl); 0.672 -> 0.713 (vd) | Helped |
| Use both directions rather than only the foreground-light direction (ablation) | BigBiGAN segmentation pipeline · Flowers | IoU: 0.479 -> 0.541 | Mixed |
| Set λ to 0.2 rather than 0.1 during latent-direction optimization (ablation) | BigBiGAN segmentation pipeline · CUB | IoU: 0.631 -> 0.658 | Helped |
| Increase λ from 0.2 to 0.4 during latent-direction optimization (ablation) | BigBiGAN segmentation pipeline · CUB | IoU: 0.658 -> 0.418 | Hurt |
| Set perturbation length ε to 2 rather than 4 (ablation) | BigBiGAN segmentation pipeline · Flowers | IoU: 0.435 -> 0.479 | Helped |
| Increase perturbation length ε from 2 to 6 (ablation) | BigBiGAN segmentation pipeline · DUTS | IoU: 0.486 -> 0.238 | Hurt |
| Replace one trained segmentation network with an ensemble of networks trained on each latent direction (ablation) | BigBiGAN segmentation pipeline · CUB | IoU: 0.664 -> 0.650 | No clear effect |
Closest papers by meaning
- Object Segmentation Without Labels with Large-Scale Generative Models (ICML 2021 · similarity 0.91)
- Training Generative Adversarial Networks from Incomplete Observations using Factorised Discriminators (ICLR 2020 · similarity 0.89)
- GANSeg: Learning To Segment by Unsupervised Hierarchical Image Generation (CVPR 2022 · similarity 0.89)
- ILSGAN: Independent Layer Synthesis for Unsupervised Foreground-Background Segmentation (AAAI 2023 · similarity 0.89)
- Emergence of Object Segmentation in Perturbed Generative Models (NeurIPS 2019 · similarity 0.88)
- A Generative Approach at the Instance-Level for Image Segmentation Under Limited Training Data Conditions (Student Abstract) (AAAI 2025 · similarity 0.85)
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.