Research map / Architectures and efficiency / CNN design
Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks
Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Andrea Vedaldi
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 parameter-free gather-excite context modulation | ResNet-50 · ImageNet-1K val | top-1 error (%): 23.30 -> 22.14 | Improved |
| Add parameterized gather and excite operators | ResNet-50 · ImageNet-1K val | top-1 error (%): 22.00 -> 21.88 | Improved |
| Add GE-θ operators in stage 2 only (ablation) | ResNet-50 · ImageNet-1K val | top-1 error (%): 23.30 -> 23.29 | No clear effect |
| Add GE-θ operators in stage 3 only (ablation) | ResNet-50 · ImageNet-1K val | top-1 error (%): 23.30 -> 22.70 | Improved |
| Add GE-θ operators in stage 4 only (ablation) | ResNet-50 · ImageNet-1K val | top-1 error (%): 23.30 -> 22.50 | Improved |
| Add GE operators in all stages rather than a single stage (ablation) | ResNet-50 · ImageNet-1K val | top-1 error (%): 22.50 -> 22.00 | Improved |
| Increase gather extent ratio (ablation) | GE-θ⁻ on ResNet-50 · ImageNet-1K val | top-1 error (%) | Improved |
| Use global max pooling instead of average pooling (ablation) | GE-θ⁻ on ResNet-50 · ImageNet-1K val | top-1 error (%): 22.14 -> 23.65 | Hurt |
| Use average pooling instead of max pooling (ablation) | GE-θ⁻ on ResNet-50 · ImageNet-1K val | top-1 error (%): 23.10 -> 22.87; 23.65 -> 22.14 | Improved |
| Add parameterized gather-excite modules | ShuffleNet · ImageNet-1K val | top-1 error (%): 32.60 -> 31.80 | Improved |
| Add GE-θ+ modules (ablation) | ShuffleNet · ImageNet-1K val | top-1 error (%): 31.80 -> 30.12 | Improved |
| Add GE-θ+ modules | Faster R-CNN with ResNet-50 · MS COCO 2014 val | mAP: 27.3 -> 28.6 | Improved |
Closest papers by meaning
- Squeeze-and-Excitation Networks (CVPR 2018 · similarity 0.85)
- Context Encoders: Feature Learning by Inpainting (CVPR 2016 · similarity 0.85)
- Contextual Convolutional Networks (ICLR 2023 · similarity 0.85)
- An Empirical Evaluation of Current Convolutional Architectures' Ability to Manage Nuisance Location and Scale Variability (CVPR 2016 · similarity 0.85)
- Weight Excitation: Built-in Attention Mechanisms in Convolutional Neural Networks (ECCV 2020 · similarity 0.84)
- Context-aware Attentional Pooling (CAP) for Fine-grained Visual Classification (AAAI 2021 · 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.
