Research map / Detection and segmentation / Open-set and few-shot detection
Unsupervised Part Discovery via Dual Representation Alignment
Jiahao Xia, Wenjian Huang, Min Gang Xu, Jianguo Zhang, Haimin Zhang, Ziyu Sheng, Dong Xu
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 area loss (ablation) | Ours (DINO pretrained) · CUB-F | NMI: 1.8 -> 45.2 | Improved |
| add concentration loss (ablation) | Ours (DINO pretrained) · CUB-F | NMI: 5.2 -> 45.2 | Improved |
| add reconstruction loss (ablation) | Ours (DINO pretrained) · CUB-F | NMI: 15.9 -> 45.2 | Improved |
| add semantic consistency loss (ablation) | Ours (DINO pretrained) · CUB-F | NMI: 41.3 -> 45.2 | Improved |
| exchange part representations between paired images (ablation) | Ours (DINO pretrained) · CUB-F (rotation 0 deg) | NMI: 43.6 -> 45.2 | Improved |
| use DINO ImageNet pretraining for the feature encoder | Ours · CUB-F | NMI: 22.0 -> 45.2 | Improved |
| use DINO ImageNet pretraining for the feature encoder | Ours · CelebA-in-the-wild (K=8) | inter-ocular NME: 8.70% -> 10.24% | Hurt |
| use DINO ImageNet pretraining for the feature encoder | Ours · PartImageNet | NMI: 22.13 -> 27.85 | Improved |
| use MAE instead of no pretraining (ablation) | Ours (ViT-B/16) · CelebA-in-the-wild (K=8) | NMI: 50.5 -> 59.5 | Improved |
| use DINO instead of no pretraining (ablation) | Ours (ViT-B/16) · CelebA-in-the-wild (K=8) | NMI: 50.5 -> 43.7 | Hurt |
| use MoCo v3 instead of no pretraining (ablation) | Ours (ViT-B/16) · CelebA-in-the-wild (K=8) | NMI: 50.5 -> 22.0 | Hurt |
| use perceptual loss instead of MSE reconstruction loss (ablation) | Ours (no pretraining) · CelebA-in-the-wild (K=8) | inter-ocular NME: 16.70% -> 8.70% | Improved |
Closest papers by meaning
- Unsupervised Part Discovery from Contrastive Reconstruction (NeurIPS 2021 · similarity 0.91)
- Unsupervised Part Discovery via Descriptor-Based Masked Image Restoration with Optimized Constraints (ICCV 2025 · similarity 0.91)
- Unsupervised Part Learning for Visual Recognition (CVPR 2017 · similarity 0.90)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.88)
- Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations (CVPR 2024 · similarity 0.87)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.87)
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.
