Research map / Architectures and efficiency / CNN design
Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 learned patch selection rather than random patch selection, retaining the Glance Step (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 74.46% -> 75.93% | Helped |
| Use learned patch selection rather than centre-corner patch selection, retaining the Glance Step (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 75.12% -> 75.93% | Helped |
| Add a separate global encoder instead of using a single encoder (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 74.26% -> 75.93% | Helped |
| Add the CNN regularization term L_cls' (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 73.94% -> 75.93% | Helped |
| Fine-tune the global encoder on H'×W' images before initialization (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 75.56% -> 75.93% | Helped |
| Use the downsampled full image for the Glance Step instead of a centre crop (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 74.70% -> 75.93% | Helped |
| Add training Stage III to fine-tune the encoders and classifier with the learned policy fixed (ablation) | GFNet (ResNet-50) · ImageNet, fixed input sequence length t=5 | Top-1 accuracy: 75.36% -> 75.93% | Helped |
| Use GFNet instead of vanilla MobileNet-V3 at the same computation budget | MobileNet-V3 · ImageNet, budgeted batch classification | Top-1 accuracy: ~2% | Helped |
| Use GFNet instead of MSDNet | DenseNet-121-based anytime classifier · ImageNet, anytime prediction, budgets from 5×10^8 to 2.2×10^9 Multiply-Adds | Top-1 accuracy difference: ~4-10% | Helped |
| Use GFNet instead of MobileNet-V3-Large baseline | MobileNet-V3-Large · ImageNet test, iPhone XS Max, 75.4% test accuracy | Inference latency: 16.3ms -> 12.7ms; 22% | Helped |
| Increase patch size (ablation) | GFNet (ResNet-50) · ImageNet, anytime prediction | Top-1 accuracy | Mixed |
| Vary maximum input sequence length T while holding computation fixed (ablation) | GFNet (ResNet-50) · ImageNet, anytime prediction | Top-1 accuracy | No clear effect |
Closest papers by meaning
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.89)
- Multi-Scale Dense Networks for Resource Efficient Image Classification (ICLR 2018 · similarity 0.88)
- An Empirical Evaluation of Current Convolutional Architectures' Ability to Manage Nuisance Location and Scale Variability (CVPR 2016 · similarity 0.88)
- Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference (CVPR 2020 · similarity 0.87)
- Dilated Residual Networks (CVPR 2017 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
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.