Research map / Adversarial robustness and security / Certified and empirical robustness
PartImageNet++ Dataset: Scaling up Part-based Models for Robust Recognition
Xiao Li, Yining Liu, Na Dong, Sitian Qin, Xiaolin Hu
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 multi-scale part supervision (MPM) to adversarially trained ResNet-50-Gelu | ResNet-50-Gelu with AT · ImageNet-1K val, average over four attacks | average robust accuracy: 19.3 -> 20.8 | Improved |
| add MPM part supervision without AT | ResNet-50-Gelu · ImageNet-C | average accuracy: 40.8 -> 42.1 | Improved |
| add MPM part supervision with AT | ResNet-50-Gelu AT · OOD datasets average | average accuracy: 24.6 -> 25.2 | Improved |
| add MPM part supervision without AT | ResNet-50-Gelu · OOD datasets average | average accuracy: 23.3 -> 23.6 | No clear effect |
| add pseudo part labels for unannotated images (ablation) | MPM · ImageNet-1K val | l_inf robust accuracy: 33.0 -> 39.1 | Improved |
| add category filter in pseudo-label generation (ablation) | Mask R-CNN pseudo labels · PIN++ part segmentation val | AP: 37.2 -> 40.7 | Improved |
| replace two-stage part model with MPM bypass layers (ablation) | part-based model · ImageNet-1K val | l_inf robust accuracy: 38.7 -> 39.1 | Improved |
| supervise three scales instead of one (ablation) | MPM · ImageNet-1K val | l_inf (8/255) robust accuracy: 8.8 -> 13.6 | Improved |
| add top-down connections (ablation) | MPM SV3 · ImageNet-1K val | l_inf (8/255) robust accuracy: 11.7 -> 13.6 | Improved |
| use part masks instead of object masks for supervision (ablation) | MPM · ImageNet-1K val | average accuracy: 22.3 -> 23.2 | Improved |
| vary lambda above zero (ablation) | MPM · Tiny-PartImageNet++ | adversarial robustness | No clear effect |
| initialize detector backbone with MPM instead of vanilla | Faster R-CNN · MS-COCO | clean AP50: 49.8 -> 50.3 | Improved |
Closest papers by meaning
- Part-Based Models Improve Adversarial Robustness (ICLR 2023 · similarity 0.93)
- "PartImageNet: A Large, High-Quality Dataset of Parts" (ECCV 2022 · similarity 0.91)
- Unsupervised Part Learning for Visual Recognition (CVPR 2017 · similarity 0.89)
- PartField: Learning 3D Feature Fields for Part Segmentation and Beyond (ICCV 2025 · similarity 0.88)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.87)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.86)
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.
