Research map / Label-efficient and robust learning / Recognition and long-tail
On the Reliability of Cue Conflict and Beyond
Pum Jun Kim, Seung-Ah Lee, Seongho Park, Dongyoon Han, Jaejun Yoo
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 |
|---|---|---|---|
| evaluate with REFINED-BIAS instead of cue-conflict stylized images | shape-augmented ResNet-50 · ResNet-50 Shape Aug | t-test p-value on preference: p=0.181 -> p=0.018 | Improved |
| evaluate with REFINED-BIAS instead of cue-conflict | contrastive ResNet-50 · ResNet-50 Contrastive | t-test p-value on preference: p=0.684 -> p=0.009 | Improved |
| evaluate with REFINED-BIAS instead of cue-conflict | adversarially trained ResNet-50 · ResNet-50 Adversarial | t-test p-value on preference: p=5.19e-05 -> p=0.489 | Improved |
| evaluate with REFINED-BIAS instead of cue-conflict | mixed-augmentation ResNet-50 · ResNet-50 Mixed Aug | preference direction: p=2.72e-04 -> p=0.002 | Improved |
| use ambiguous non-distinctive classes in REFINED-BIAS (ablation) | Shape Aug ResNet-50 · ResNet-50 Shape Aug | t-test p-value on preference: p=0.018 -> p=0.919 | Hurt |
| use ambiguous non-distinctive classes in REFINED-BIAS (ablation) | Adversarial ResNet-50 · ResNet-50 Adversarial | t-test p-value on preference: p=0.489 -> p=0.006 | Hurt |
| replace stylized cues with pure curated cues (human agreement) | texture cues · Texture cues | Fleiss kappa: 0.2937 -> 0.7973 | Improved |
| replace stylized cues with pure curated cues (shape) | shape cues · Shape cues | Fleiss kappa: 0.7276 -> 0.9836 | Improved |
| use REFINED-BIAS cues for CNN recognition | ImageNet-1k CNNs · shape cues | mean top-1 accuracy: 4 -> 46 | Improved |
| remove human curation from dataset generation (ablation) | REFINED-BIAS · learning strategies | shape preference significance | No clear effect |
| use Swin or CMT instead of ViT | ViT · REFINED-BIAS | shape sensitivity | Improved |
| increase quantization strength from 32-bit to 2-bit | ResNets · REFINED-BIAS | shape and texture sensitivity | Hurt |
Closest papers by meaning
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.88)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.88)
- Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision (NeurIPS 2026 · similarity 0.86)
- Evaluating Multiview Object Consistency in Humans and Image Models (NeurIPS 2024 · similarity 0.86)
- Contour Integration Underlies Human-Like Vision (ICML 2025 · similarity 0.85)
- Cue3D: Quantifying the Role of Image Cues in Single-Image 3D Generation (NeurIPS 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.
