Research map / Label-efficient and robust learning / Domain adaptation and generalization
Attention Bridging Network for Knowledge Transfer
Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, Yun Fu
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 source-domain data with classification loss only (ablation) | Grad-CAM baseline · VOC 2012 val (direct attention-map evaluation) | mIoU: 30.2 -> 31.7 | No clear effect |
| add attention constraint loss with fine-tuning (ablation) | Grad-CAM + source data · VOC 2012 val (direct attention-map evaluation) | mIoU: 31.7 -> 39.4 | Improved |
| use joint two-domain training with weight sharing instead of fine-tuning (ablation) | attention constraint model · VOC 2012 val (direct attention-map evaluation) | mIoU: 39.4 -> 46.2 | Improved |
| add adversary training loss (ablation) | joint training with attention constraint · VOC 2012 val (direct attention-map evaluation) | mIoU: 46.2 -> 50.9 | Improved |
| add adversary training loss (ablation) | joint training with attention constraint · VOC 2012 test (direct attention-map evaluation) | mIoU: 47.5 -> 51.2 | Improved |
| increase source domain data from 2K to 20K images (ablation) | AttnBN · VOC 2012 val (direct attention-map evaluation) | mIoU: 45.1 -> 50.9 | Improved |
| use only 10% source data (about 2K images) (ablation) | AttnBN vs Grad-CAM · VOC 2012 val (direct attention-map evaluation) | mIoU: 14.9% | Improved |
| use AttnBN attention maps as foreground cues | DSRG · VOC 2012 val (weakly-supervised segmentation) | mIoU: 59.0 -> 62.1 | Improved |
| use AttnBN attention maps as foreground cues | DSRG · VOC 2012 test (weakly-supervised segmentation) | mIoU: 60.4 -> 63.0 | Improved |
| add attention constraint loss for domain generalization (ablation) | Ls-cl classifier · Single-label to multi-label domain | mAP: 61.3 -> 66.7 | Improved |
| add attention constraint loss to adversarial adaptation (ablation) | Ls-cl + Ladv · Single-label to multi-label domain | mAP: 64.5 -> 71.0 | Improved |
| add attention constraint loss for generalization (ablation) | Ls-cl classifier · Caltech-101 to VOC 2007 | accuracy: 0.62 -> 0.71 | Improved |
Closest papers by meaning
- Guided Attention Inference Network (TPAMI 2019 · similarity 0.89)
- Tell Me Where to Look: Guided Attention Inference Network (CVPR 2018 · similarity 0.89)
- Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer (ICLR 2017 · similarity 0.88)
- Heterogeneous Graph Attention Network for Unsupervised Multiple-Target Domain Adaptation (TPAMI 2020 · similarity 0.87)
- Deep Adversarial Attention Alignment for Unsupervised Domain Adaptation: the Benefit of Target Expectation Maximization (ECCV 2018 · similarity 0.87)
- L2G: A Simple Local-to-Global Knowledge Transfer Framework for Weakly Supervised Semantic Segmentation (CVPR 2022 · similarity 0.87)
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.
