Research map / Architectures and efficiency / Attention and transformers
PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers
Ananthu Aniraj, Cassio F. Dantas, Dino Ienco, Diego Marcos
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 |
|---|---|---|---|
| replace concentration prior with total variation prior on a self-supervised ViT backbone | PDiscoNet vs PDiscoFormer · CUB (K=16) | ARI: 43.41 -> 55.83 | Improved |
| replace concentration prior with total variation prior on irregular multi-instance parts | PDiscoNet + ViT-B vs PDiscoFormer · Oxford Flowers (K=8) | foreground mIoU: 13.18 -> 69.59 | Improved |
| add total variation loss (ablation) | PDiscoFormer · PartImageNet OOD (K=25) | ARI: 25.35 -> 59.27 | Improved |
| add total variation loss (ablation) | PDiscoFormer · CUB (K=16) | Top-1 accuracy: 80.22 -> 88.72 | Improved |
| add equivariance loss (ablation) | PDiscoFormer · CUB (K=16) | ARI: 30.63 -> 55.83 | Improved |
| add per-part layer normalization modulation (ablation) | PDiscoFormer · CUB (K=16) | Top-1 accuracy: 69.73 -> 88.72 | Improved |
| add presence loss L_p1 (each part present at least once per batch) (ablation) | PDiscoFormer · CUB (K=16) | ARI: 34.28 -> 55.83 | Improved |
| add background presence loss L_p0 (ablation) | PDiscoFormer · PartImageNet OOD (K=25) | ARI: 49.08 -> 59.27 | Improved |
| add Gumbel-Softmax instead of Softmax on part maps (ablation) | PDiscoFormer · CUB (K=16) | ARI: 43.61 -> 55.83 | Improved |
| add orthogonality loss (ablation) | PDiscoFormer · PartImageNet OOD (K=25) | ARI: 58.86 -> 59.27 | No clear effect |
| add entropy loss (ablation) | PDiscoFormer · PartImageNet OOD (K=25) | Top-1 accuracy: 91.19 -> 90.77 accuracy; NMI 39.16 -> 44.71 | Mixed |
| use DINOv2 ViT-B instead of ResNet-101 backbone with the TV-based losses (ablation) | PDiscoFormer · Oxford Flowers | Top-1 accuracy: 8.41 -> 99.61 (K=2) | Improved |
Closest papers by meaning
- PDiscoNet: Semantically consistent part discovery for fine-grained recognition (ICCV 2023 · similarity 0.90)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.86)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 2025 · similarity 0.86)
- Activating Sparse Part Concepts for 3D Class Incremental Learning (CVPR 2025 · similarity 0.85)
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.85)
- Vision Transformers Need Registers (ICLR 2024 · 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.
