Research map / Detection and segmentation / CNN features
Attend and Rectify: a gated attention mechanism for fine-grained recovery
Pau Rodriguez, Josep M. Gonfaus, Guillem Cucurull, F. XavierRoca, Jordi Gonzalez
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 |
|---|---|---|---|
| augment WRN-28 with attention (WARN, no dropout, AD2 AW4) | WRN-28 · CIFAR-10 (from scratch) | error rate (%): 4 -> 3.60 | Improved |
| augment WRN-28 with attention (WARN, AD3 AW4, dropout) | WRN-28 · CIFAR-100 (from scratch) | error rate (%): 18.85 -> 18.26 | Improved |
| augment WRN-40 with attention (WARN) | WRN-40 · CIFAR-100 (from scratch) | error rate (%): 18.3 -> 17.82 | Improved |
| use WARN attention instead of WRN-att2 | WRN-40 · CIFAR-100 (from scratch) | error rate (%): 19.20 -> 17.82 | Improved |
| increase attention depth from 1 to number of residual blocks and width from 1 to 4 (ablation) | WARN · CIFAR-100 (from scratch) | error rate (%): 1.5% accuracy | Improved |
| use gating instead of linear projection (ablation) | WARN · CIFAR-100 (from scratch) | accuracy: 4% better | Improved |
| use gating instead of averaging output vectors (ablation) | WARN · CIFAR-100 (from scratch) | accuracy: 3% better | Improved |
| add regularization of attention heads (ablation) | WARN · CIFAR-100 (from scratch) | accuracy: 0.6% improvement | Improved |
| use sigmoid attention instead of softmax (ablation) | WARN · CIFAR-100 (from scratch) | accuracy | No clear effect |
| allow gradients from attention modules to the base network (ablation) | WARN · CIFAR-10 (from scratch) | accuracy: 0.2% drop when blocked | Improved |
| add gates, softmax and regularization (best configuration) (ablation) | CNN + attention · Cluttered Translated MNIST | accuracy (%): 97.8 | Improved |
| add attention to ImageNet-pretrained WRN-50-4 | WRN-50-4 · Stanford Dogs (fine-tuned) | accuracy (%): 89.6 -> 92.9 | Improved |
Closest papers by meaning
- Looking for the Devil in the Details: Learning Trilinear Attention Sampling Network for Fine-Grained Image Recognition (CVPR 2019 · similarity 0.87)
- Improving Fine-Grained Visual Recognition in Low Data Regimes via Self-Boosting Attention Mechanism (ECCV 2022 · similarity 0.87)
- Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-Grained Image Recognition (CVPR 2017 · similarity 0.87)
- Dual Attention Networks for Few-Shot Fine-Grained Recognition (AAAI 2022 · similarity 0.87)
- Learning a Discriminative Filter Bank Within a CNN for Fine-Grained Recognition (CVPR 2018 · similarity 0.86)
- Multi-Attention Multi-Class Constraint for Fine-grained Image Recognition (ECCV 2018 · similarity 0.86)
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.
