Research map / Detection and segmentation / CNN features
Attention to Scale: Scale-Aware Semantic Image Segmentation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use learned scale attention rather than average pooling (ablation) | DeepLab-LargeFOV with multi-scale inputs · PASCAL VOC 2012 validation, ImageNet pretrained, 3 scales with extra supervision | mIOU: 67.98 -> 69.08 | Helped |
| Use learned scale attention rather than max pooling (ablation) | DeepLab-LargeFOV with multi-scale inputs · PASCAL VOC 2012 validation, ImageNet pretrained, 3 scales with extra supervision | mIOU: 67.79 -> 69.08 | Helped |
| Add supervision to each scale's output (ablation) | DeepLab-LargeFOV with multi-scale feature merging · PASCAL-Person-Part validation | mIOU: 53.49 -> 55.85 | Helped |
| Add supervision to each scale's output (ablation) | DeepLab-LargeFOV with multi-scale feature merging · PASCAL VOC 2012 validation, ImageNet pretrained | mIOU: 64.37 -> 69.08 | Helped |
| Add supervision to each scale's output (ablation) | DeepLab-LargeFOV with multi-scale feature merging · MS-COCO 2014 subset validation | mIOU: 33.42 -> 35.78 | Helped |
| Use 3 input scales instead of 2 with attention and extra supervision (ablation) | DeepLab-LargeFOV · PASCAL VOC 2012 validation, ImageNet pretrained | mIOU: 68.24 -> 69.08 | Helped |
| Use 3 input scales instead of 2 with attention and extra supervision (ablation) | DeepLab-LargeFOV · PASCAL-Person-Part validation | mIOU: 55.85 -> 56.39 | Helped |
| Use 3 input scales instead of 2 with attention and extra supervision (ablation) | DeepLab-LargeFOV · MS-COCO 2014 subset validation | mIOU: 35.41 -> 35.78 | Helped |
| Add a fourth input scale at 0.25 (ablation) | Three-scale attention model · PASCAL-Person-Part validation | mIOU: 0.5% | Hurt |
| Use fc_8 rather than fc_7 features in the attention model (ablation) | Attention model · PASCAL-Person-Part validation | mIOU: ~0.5% | Hurt |
| Vary the attention network depth, first-layer kernel size, or filter count (ablation) | Attention model · PASCAL-Person-Part validation | mIOU: 0.1 % to 0.4 % | No clear effect |
| Add random input scaling from 0.6 to 1.4 during training (ablation) | DeepLab-CRF-COCO-LargeFOV-Attention · PASCAL VOC 2012 test | mIOU: 0.6% | Helped |
Closest papers by meaning
- Hierarchically Gated Deep Networks for Semantic Segmentation (CVPR 2016 · similarity 0.90)
- Squeeze-and-Attention Networks for Semantic Segmentation (CVPR 2020 · similarity 0.90)
- Multi-Scale Representations by Varying Window Attention for Semantic Segmentation (ICLR 2024 · similarity 0.89)
- Fully Convolutional Neural Networks with Full-Scale-Features for Semantic Segmentation (AAAI 2017 · similarity 0.88)
- Rethinking Alignment and Uniformity in Unsupervised Image Semantic Segmentation (AAAI 2023 · similarity 0.88)
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · 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.