Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Attributes-Guided and Pure-Visual Attention Alignment for Few-Shot Recognition
Siteng Huang, Min Zhang, Yachen Kang, Donglin Wang
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 AGAM to ProtoNet | ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 53.01 ± 0.34 -> 75.87 ± 0.29; 71.91 ± 0.22 -> 81.66 ± 0.25 | Improved |
| Add AGAM to ProtoNet | ProtoNet, Conv-4 · SUN 5-way 1-shot and 5-shot | accuracy: 57.76 ± 0.29 -> 65.15 ± 0.31; 79.27 ± 0.19 -> 80.08 ± 0.21 | Improved |
| Add human-annotated attributes to AGAM (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 67.35 ± 0.32 -> 75.87 ± 0.29; 77.87 ± 0.19 -> 81.66 ± 0.25 | Improved |
| Replace real attributes with all-0 attributes (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 75.87 ± 0.29 -> 63.47 ± 0.35; 81.66 ± 0.25 -> 75.27 ± 0.20 | Hurt |
| Swap spatial and channel attention order (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 75.87 ± 0.29 -> 74.22 ± 0.27; 81.66 ± 0.25 -> 79.72 ± 0.26 | Hurt |
| Remove average pooling (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 75.87 ± 0.29 -> 66.27 ± 0.29; 81.66 ± 0.25 -> 76.58 ± 0.25 | Hurt |
| Remove max pooling (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 75.87 ± 0.29 -> 67.60 ± 0.29; 81.66 ± 0.25 -> 77.09 ± 0.22 | Hurt |
| Remove channel-wise attention (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot | accuracy: 75.87 ± 0.29 -> 54.91 ± 0.36 | Hurt |
| Remove spatial-wise attention (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 75.87 ± 0.29 -> 69.66 ± 0.31; 81.66 ± 0.25 -> 76.24 ± 0.27 | Hurt |
| Add both channel and spatial attention alignment losses (ablation) | AGAM with ProtoNet, Conv-4 · CUB 5-way 1-shot and 5-shot | accuracy: 75.37 ± 0.31 -> 75.87 ± 0.29; 78.92 ± 0.27 -> 81.66 ± 0.25 | Improved |
| Use soft margin instead of L1, MSE, or smoothL1 for attention alignment (ablation) | AGAM · CUB and SUN 5-way 1-shot and 5-shot | accuracy | Improved |
Closest papers by meaning
- Fine-Grained Generalized Zero-Shot Learning via Dense Attribute-Based Attention (CVPR 2020 · similarity 0.89)
- CAD: Co-Adapting Discriminative Features for Improved Few-Shot Classification (CVPR 2022 · similarity 0.88)
- SAGE: Structured Attribute-Guided Enhancement for GZSL (AAAI 2026 · similarity 0.88)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.86)
- Class Semantic Attribute Perception Guided Zero-Shot Learning (AAAI 2025 · similarity 0.86)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · 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.
