Research map / Detection and segmentation / CNN features
Just Noticeable Difference Modeling for Deep Visual Features
Rui Zhao, Wenrui Li, Lin Zhu, Yajing Zheng, Weisi Lin
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 |
|---|---|---|---|
| use FeatJND-shaped distortion instead of Gaussian noise at matched NRMSE | feature-space distortion · ImageNet-1K, COCO (all backbones) | task performance | Improved |
| use FeatJND-guided token-wise step sizes instead of random permutation of step sizes | token-wise feature quantization · ImageNet / COCO, transformer backbones | accuracy / mAP | Improved |
| use FeatJND-guided quantization instead of global uniform step | token-wise feature quantization · ImageNet / COCO, transformer backbones | accuracy / mAP | Improved |
| increase loss weight lambda_t from 10 to 50 (ablation) | FeatJND · Swin-S ImageNet-1K, NRMSE 40 | Top-1 accuracy (%): 53.7 -> 82.5 | Improved |
| increase lambda_t from 10 to 200 (ablation) | FeatJND · Swin-S + Mask R-CNN detection, NRMSE 5 | mAP (%): 38.3 -> 47.5 | Improved |
| increase lambda_t beyond 50 (50 to 70) (ablation) | FeatJND · Swin-S ImageNet-1K | Top-1 accuracy (%): 82.5 -> 82.5 | No clear effect |
| raise distillation temperature T from 1 to 4 (ablation) | FeatJND · Swin-S ImageNet-1K, NRMSE 40 | Top-1 accuracy (%): 78.7 -> 82.5 | Improved |
| raise temperature T from 4 to 8 (ablation) | FeatJND · Swin-S ImageNet-1K | Top-1 accuracy (%): 82.5 -> 82.4 | No clear effect |
| raise temperature T from 1 to 4 (ablation) | FeatJND · Swin-S + Mask R-CNN detection, NRMSE 5 | mAP (%): 45.1 -> 47.5 | Improved |
| vary number of residual blocks in estimator (ablation) | FeatJND · Swin-S ImageNet-1K | Top-1 accuracy (%) | No clear effect |
| use larger 5x5 kernels instead of 1x1 in estimator (ablation) | FeatJND · Swin-S ImageNet-1K, NRMSE 40 | Top-1 accuracy (%): 82.5 -> 81.8 | Hurt |
| scale FeatJND strength alpha above 1 (ablation) | FeatJND · ImageNet / COCO | performance drop | Hurt |
Closest papers by meaning
- FeatUp: A Model-Agnostic Framework for Features at Any Resolution (ICLR 2024 · similarity 0.84)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.83)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.83)
- VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models (ICML 2026 · similarity 0.83)
- Test-Time Canonicalization by Foundation Models for Robust Perception (ICML 2025 · similarity 0.83)
- OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models (CVPR 2026 · similarity 0.83)
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.
