Research map / Detection and segmentation / CNN features
TDAM: Top-Down Attention Module for Contextually Guided Feature Selection in CNNs
Shantanu Jaiswal, Basura Fernando, Cheston Tan
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 TDjoint (t=2, m=1) top-down attention module | ResNet50 · ImageNet-V1 val | top-1 accuracy: 77.51 -> 78.96 | Improved |
| add TDjoint (t=2, m=1) top-down attention module | ResNet50 · ImageNet-V2 val | top-1 accuracy: 66.39 -> 67.66 | Improved |
| use TDjoint instead of FCA-TS attention | ResNet50 · ImageNet-V1 val | top-1 accuracy: 78.70 -> 78.96 | Improved |
| use TDjoint (t=2, m=1) instead of SE | ResNet101 · ImageNet-V1 val | top-1 accuracy: 80.84 -> 81.62 | Improved |
| apply TD module only at layer 4 plus alternating blocks (L4) instead of layers 3 and 4 (ablation) | ResNet101 TDjoint · ImageNet-V1 val | top-1 accuracy: 81.62 -> 81.12 | Hurt |
| increase feedback distance from m=1 to m=3 in TDtop (ablation) | ResNet50 · ImageNet-V1 val | top-1 accuracy: 78.82 -> 78.90 | No clear effect |
| add top-down attention module for test resolution 448x448 | ResNet50 · ImageNet-V1 val (448x448 test) | top-1 accuracy: 2% better than original ResNet | Improved |
| use feedback distance of at least 1 (distinct top and bottom maps) (ablation) | TD module on ResNet50 · ImageNet-V1 val | top-1 accuracy | Improved |
| use channel-then-spatial attention instead of parallel channel and spatial (ablation) | TD module on ResNet50 · ImageNet-V1 val | top-1 accuracy | Improved |
| add spatial attention to channel-only feedback (ablation) | TD module on ResNet50 · ImageNet-V1 val | top-1 accuracy | Improved |
| use top-down attention instead of recurrent feedback convolution (ablation) | TD module on ResNet50 · ImageNet-V1 val | top-1 accuracy | Improved |
| increase feedback steps beyond t=2 (ablation) | TD module · ImageNet-200 subset | top-1 accuracy | Hurt |
Closest papers by meaning
- TDAF: Top-Down Attention Framework for Vision Tasks (AAAI 2021 · similarity 0.88)
- CBAM: Convolutional Block Attention Module (ECCV 2018 · similarity 0.88)
- Top-Down Neural Attention by Excitation Backprop (ECCV 2016 · similarity 0.88)
- Learn to Pay Attention (ICLR 2018 · similarity 0.87)
- SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks (ICML 2021 · similarity 0.87)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · 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.
