Research map / Detection and segmentation / CNN features
TDAF: Top-Down Attention Framework for Vision Tasks
Bo Pang, Yizhuo Li, Jiefeng Li, Muchen Li, Hanwen Cao, Cewu Lu
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 top-down ANAR attention to VGG16 | VGG16 · CIFAR-10 test | accuracy: 92.64 -> 93.43 | Improved |
| Replace ANAR-3's learnable DeCNN up-sampling with parameter-free interpolation (ANAR-2) (ablation) | VGG16 · CIFAR-10 test | accuracy: 93.43 -> 93.02 | Hurt |
| Add top-down ANAR attention | ImageNet ResNet50 · ImageNet-2012 validation | Top-1 accuracy: 77.1 -> 78.4 | Improved |
| Increase ANAR depth from 3 to 7 layers (ablation) | ImageNet ResNet50 with 3 flows · ImageNet-2012 validation | Top-1 accuracy: 78.4 -> 78.8 | Improved |
| Increase recursive flows from 3 to 4 (ablation) | ImageNet ResNet101 with ANAR-3 · ImageNet-2012 validation | Top-1 accuracy: 79.9 -> 80.2 | Improved |
| Remove inter-flow attention and concatenate standalone flow features (ablation) | ImageNet ResNet50, 3-flow R2DNS · ImageNet-2012 validation | Top-1 accuracy: 78.6 -> 77.5 | Hurt |
| Add top-down attention to FCOS | ResNet50-FCOS · COCO-2017 | AP: 37.1 -> 39.8 | Improved |
| Increase recursive flows from 3 to 4 (ablation) | COCO ResNet101-FCOS with ANAR-3 · COCO-2017 | AP: 42.0 -> 42.3 | Improved |
| Increase ANAR depth from 3 to 5 layers (ablation) | COCO ResNet50-FCOS with 3 flows · COCO-2017 | AP: 39.6 -> 39.8 | Improved |
| Add top-down attention to SimplePose ResNet50 | SimplePose ResNet50 · COCO-2017 validation, pose (GT detector) | AP: 72.7 -> 74.3 | Improved |
| Increase ANAR depth from 3 to 7 layers (ablation) | SimplePose ResNet50 with 3 flows · COCO-2017 validation, pose (GT detector) | AP: 73.8 -> 74.3 | Improved |
| Add TDAF attention | 3D-ResNet50, pre-trained · Kinetics-400 validation | accuracy: 68.9 -> 70.1 | Improved |
Closest papers by meaning
- TDAM: Top-Down Attention Module for Contextually Guided Feature Selection in CNNs (ECCV 2022 · similarity 0.88)
- TdAttenMix: Top-Down Attention Guided Mixup (AAAI 2025 · similarity 0.84)
- Top-Down Visual Attention From Analysis by Synthesis (CVPR 2023 · similarity 0.84)
- Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers (CVPR 2024 · similarity 0.83)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.83)
- OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels (CVPR 2025 · similarity 0.83)
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.
