Research map / Architectures and efficiency / RNNs and normalization
Resolution Switchable Networks for Runtime Efficient Image Recognition
Yikai Wang, Fuchun Sun, Duo Li, Anbang Yao
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-resolution ensemble distillation (MRED) to parallel training (ablation) | RS-Net · ImageNet val 96x96 (ResNet18) | top-1 accuracy: 61.5 -> 64.1 | Improved |
| parallel training of five resolutions with shared weights and private BN (no MRED) | ResNet18 · ImageNet val 96x96 | top-1 accuracy: 62.6 -> 61.5 | Hurt |
| parallel training of five resolutions (no MRED) | ResNet18 · ImageNet val 224x224 | top-1 accuracy: 71.0 -> 73.0 | Improved |
| train one RS-Net instead of an individual model at 224 only (I-224) | ResNet18 · ImageNet val 96x96 | top-1 accuracy: 42.5 -> 64.1 | Improved |
| add high-to-low resolution distillation paths (full over vanilla MRED) (ablation) | MRED ResNet18 · ImageNet val 96x96 | top-1 accuracy: 63.7 -> 64.1 | Improved |
| add distillation from p1..p(S-1) paths in addition to p0 (ablation) | MRED ResNet18 · ImageNet val 224x224 | top-1 accuracy: 72.2 -> 73.1 | Improved |
| add distillation from ensemble p0 (vs only p1..p(S-1)) (ablation) | MRED ResNet18 · ImageNet val 224x224 | top-1 accuracy: 72.3 -> 73.1 | Improved |
| use single-crop instead of multi-crop across resolutions (ablation) | RS-Net ResNet18 · ImageNet val 96x96 | top-1 accuracy: 62.8 -> 64.1 | Improved |
| use multiple resolutions instead of identical resolutions for ensemble distillation (ablation) | RS-Net ResNet18 · ImageNet val 224x224 | top-1 accuracy: 72.1 -> 73.1 | Improved |
| train with area ratio [0.3,1.0] instead of [0.08,1.0] to reduce train-test discrepancy (ablation) | parallel training ResNet18 · ImageNet val | top-1 accuracy: interaction effects weakened | Mixed |
| adjust resolution instead of width at comparable FLOPs | MobileNetV2 I-Nets · ImageNet val 96x96 vs 0.35x width | top-1 accuracy: 60.1 -> 62.6 | Improved |
| apply RS-Net training to quantized ResNet50 (2/2 bits) | LQ-Nets ResNet50 · ImageNet val 224x224 | top-1 accuracy: 72.2 -> 74.0 | Improved |
Closest papers by meaning
- Dynamic Resolution Network (NeurIPS 2021 · similarity 0.87)
- Learning to Learn Parameterized Classification Networks for Scalable Input Images (ECCV 2020 · similarity 0.86)
- MTKD: Multi-Teacher Knowledge Distillation for Image Super-Resolution (ECCV 2024 · similarity 0.86)
- Learning with Privileged Information for Efficient Image Super-Resolution (ECCV 2020 · similarity 0.85)
- Resolution Adaptive Networks for Efficient Inference (CVPR 2020 · similarity 0.85)
- Rethinking Resolution in the Context of Efficient Video Recognition (NeurIPS 2022 · 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.
