Research map / Vision-language and multimodal / Human and embodied modeling
Exploring Simple and Transferable Recognition-Aware Image Processing
Zhuang Liu, Hung-Ju Wang, Tinghui Zhou, Zhiqiang Shen, Bingyi Kang, Evan Shelhamer, Trevor Darrell
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 recognition loss from a fixed pretrained recognizer | SRResNet super-resolution · ImageNet val (ResNet-18) | top-1 accuracy: 52.6 -> 61.8 | Improved |
| add recognition loss, JPEG-deblocking | SRResNet · ImageNet val (ResNet-18) | top-1 accuracy: 48.2 -> 57.7 | Improved |
| add recognition loss, denoising | SRResNet · ImageNet val (ResNet-18) | top-1 accuracy: 61.9 -> 65.1 | Improved |
| add an intermediate transformer between processor and recognizer | RA processing, super-resolution · ImageNet val (ResNet-18) | top-1 accuracy: 61.8 -> 63.0 | Improved |
| use unsupervised RA (distill recognizer output) instead of label-based RA | RA processing, super-resolution · ImageNet val (ResNet-18) | top-1 accuracy: 61.8 -> 61.3 | Hurt |
| train the processor with recognizer ResNet-18 and evaluate on ResNet-50 | RA processing, super-resolution · ImageNet val (ResNet-50) | top-1 accuracy: 58.8 -> 66.7 | Improved |
| transformer variant trained on ResNet-18, evaluated on VGG-16 | RA w/ Transformer, super-resolution · ImageNet val (VGG-16) | top-1 accuracy: 50.2 -> 27.0 | Hurt |
| add RA loss when transferring across disjoint category splits | RA processing, super-resolution · ImageNet Cat A to Cat B | 500-way accuracy: 60.1 -> 66.5 | Improved |
| add RA loss on randomized compound corruption | RA processing · ImageNet val (ResNet-18) | top-1 accuracy: 31.5 -> 41.2 | Improved |
| set lambda to 1e-3 instead of 0 for image quality tradeoff (ablation) | RA processing, super-resolution · ImageNet SR PSNR | PSNR: 26.73 -> 26.31 | Hurt |
| raise lambda from 1e-3 to 1e-2 (ablation) | RA processing, super-resolution · ImageNet val (ResNet-18) | top-1 accuracy: 61.8 -> 61.3 | Hurt |
| add the image processing loss alongside the recognition loss (ablation) | RA processing, super-resolution · ImageNet val (ResNet-18) | top-1 accuracy: 60.9 -> 61.8 | Improved |
Closest papers by meaning
- Revisiting Classifier: Transferring Vision-Language Models for Video Recognition (AAAI 2023 · similarity 0.86)
- Bag of Tricks for Image Classification with Convolutional Neural Networks (CVPR 2019 · similarity 0.86)
- Do Better ImageNet Models Transfer Better? (CVPR 2019 · similarity 0.86)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · similarity 0.84)
- ADDP: Learning General Representations for Image Recognition and Generation with Alternating Denoising Diffusion Process (ICLR 2024 · similarity 0.84)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.84)
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.
