Research map / Detection and segmentation / CNN features
Expectation-Maximization Attention Networks for Semantic Segmentation
Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, Hong Liu
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 |
|---|---|---|---|
| update initial bases by moving average instead of back-propagation or no update (ablation) | EMAU · PASCAL VOC val (R50) | mIoU: surpasses others by at least 0.9 in mIoU | Improved |
| update initial bases by back-propagation instead of keeping them fixed (ablation) | EMAU · PASCAL VOC val (R50) | mIoU | Hurt |
| apply L2Norm to bases instead of no normalization (ablation) | EMAU · PASCAL VOC val (R50) | mIoU | Improved |
| apply LayerNorm to bases instead of no normalization (ablation) | EMAU · PASCAL VOC val (R50) | mIoU | Hurt |
| increase evaluation iterations T_eval (ablation) | EMAU · PASCAL VOC val (R50) | mIoU | Improved |
| increase training iterations T_train beyond 3 (ablation) | EMAU · PASCAL VOC val (R50) | mIoU | Hurt |
| EM iterations (3) versus single iteration A2 block (ablation) | EMANet · PASCAL VOC val (R50) | mIoU: 77.41 vs 77.34 | No clear effect |
| EM iterations versus non-local module (ablation) | EMANet · PASCAL VOC val (R50) | mIoU: 77.78 vs 77.34 | No clear effect |
| widen EMANet input channels from 256 to 512 | EMANet · PASCAL VOC val SS | mIoU: 79.73 -> 80.05 | Improved |
| replace DeeplabV3 with EMANet (512), single scale | ResNet-101 · PASCAL VOC val SS | mIoU: 78.51 -> 80.05 | Improved |
| replace DeeplabV3 with EMANet (512), MS+Flip | ResNet-101 · PASCAL VOC val MS+Flip | mIoU: 79.77 -> 81.32 | Improved |
| replace DeeplabV3 by EMANet on test | ResNet-101 · PASCAL VOC test | mIoU: 85.7 -> 87.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- DANet — DANet: Divergent Activation for Weakly Supervised Object Localization (ICCV 2019) · better on 4 of 4 shared results
- PSANet — PSANet: Point-wise Spatial Attention Network for Scene Parsing (ECCV 2018) · better on 3 of 3 shared results
- Exfuse — ExFuse: Enhancing Feature Fusion for Semantic Segmentation (ECCV 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Squeeze-and-Attention Networks for Semantic Segmentation (CVPR 2020 · similarity 0.87)
- Explicitly Modeled Attention Maps for Image Classification (AAAI 2021 · similarity 0.87)
- Online Attention Accumulation for Weakly Supervised Semantic Segmentation (TPAMI 2021 · similarity 0.87)
- Attention to Scale: Scale-Aware Semantic Image Segmentation (CVPR 2016 · similarity 0.87)
- Attention (as Discrete-Time Markov) Chains (NeurIPS 2025 · similarity 0.86)
- 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.
