Research map / Label-efficient and robust learning / Recognition and long-tail
Mitigating the Effect of Incidental Correlations on Part-based Learning
Gaurav Bhatt, Deepayan Das, Leonid Sigal, Vineeth N. Balasubramanian
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 foreground-background mixture regularization L_mix (ablation) | DPViT pretraining · MiniImageNet few-shot learning | 1-shot accuracy: 61.24 -> 62.81 | Improved |
| add foreground-background mixture regularization L_mix (ablation) | DPViT pretraining · MiniImageNet few-shot learning | 5-shot accuracy: 81.12 -> 83.25 | Improved |
| add sparse and orthogonal quality regularization L_Q (ablation) | DPViT pretraining · MiniImageNet few-shot learning | 1-shot accuracy: 62.81 -> 62.15 | Hurt |
| add sparse and orthogonal quality regularization L_Q (ablation) | DPViT pretraining · MiniImageNet part representations | ||P||_1: 8.73 -> 0.35 | Improved |
| add classification-token invariance loss L_cls^inv (ablation) | DPViT fine-tuning · ImageNet-9 M-RAND | accuracy: 82.7 -> 87.5 | Improved |
| add classification-token invariance loss L_cls^inv (ablation) | DPViT fine-tuning · ImageNet-9 BG-GAP | BG-GAP: 8.8 -> 5.9 | Improved |
| add foreground-code invariance loss L_p^inv (ablation) | DPViT fine-tuning with L_cls^inv · ImageNet-9 M-RAND | accuracy: 87.5 -> 87.2 | Hurt |
| increase λ_s and λ_o from 0.5 to 2.0 (ablation) | DPViT pretraining · MiniImageNet | 1-shot accuracy: 73.8 -> 72.4 | Hurt |
| use foreground masks for only 10% of training samples instead of all samples (ablation) | DPViT pretraining · MiniImageNet | 1-shot accuracy: 73.8 -> 72.9 | No clear effect |
| use K=32 instead of K=64 with n_f=K/2 (ablation) | DPViT · MiniImageNet | 1-shot accuracy: 72.9 -> 72.2 | Hurt |
| share MSA and MCA attention layers (ablation) | DPViT · MiniImageNet | 1-shot accuracy: 73.8 -> 73.6 | No clear effect |
| replace L2 mix loss with L1 (ablation) | DPViT pretraining · MiniImageNet | 1-shot accuracy: 65.12 -> 58.73 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FeLMi — FeLMi : Few shot Learning with hard Mixup (NeurIPS 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unsupervised Part Discovery via Dual Representation Alignment (TPAMI 2024 · similarity 0.88)
- Unsupervised Part Discovery from Contrastive Reconstruction (NeurIPS 2021 · similarity 0.87)
- "PartImageNet: A Large, High-Quality Dataset of Parts" (ECCV 2022 · similarity 0.87)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.87)
- PARTONOMY: Large Multimodal Models with Part-Level Visual Understanding (NeurIPS 2025 · similarity 0.85)
- Learning Partonomic 3D Reconstruction from Image Collections (CVPR 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.
