Research map / Detection and segmentation / CNN features
Region-Based Semantic Segmentation with End-to-End Training
Holger Caesar, Jasper Uijlings, Vittorio Ferrari
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 region-to-pixel layer with end-to-end training (ablation) | Region-based baseline with bounding box pooling · SIFT Flow test | global accuracy: 60.9 -> 83.7 | Improved |
| apply max before softmax (ablation) | Softmax-before-max variant · SIFT Flow test | class-average accuracy: +8.7% | Improved |
| use multi-scale overlapping regions instead of non-overlapping (ablation) | Non-overlapping regions · SIFT Flow test | class-average accuracy: 60.0 -> 64.0 | Improved |
| use free-form ROI pooling instead of bounding box pooling (ablation) | Bounding box ROI pooling · SIFT Flow test | class-average accuracy: 62.3 -> 62.8 | Improved |
| combine free-form and bounding box pooling with tied weights (ablation) | Free-form ROI pooling · SIFT Flow test | class-average accuracy: 62.8 -> 63.4 | Improved |
| use separate weights for the two pooling streams (ablation) | Tied weights combination · SIFT Flow test | class-average accuracy: 63.4 -> 64.0 | Improved |
| use free-form regions instead of purely rectangular regions (ablation) | Purely rectangular regions · SIFT Flow test | class-average accuracy: 59.3 -> 64.0 | Improved |
| use region-based end-to-end model instead of FCN-16s on boundaries | FCN-16s · SIFT Flow test, boundary pixels | class-average accuracy: 37.9 -> 57.3 | Improved |
| use region-based end-to-end model instead of FCN-16s on full image | FCN-16s · SIFT Flow test, full image | class-average accuracy: 49.3 -> 64.0 | Improved |
| use region-based end-to-end model instead of FCN-16s on boundaries | FCN-16s · PASCAL Context val, boundary pixels | class-average accuracy: 34.0 -> 38.9 | Improved |
| use unbalanced loss model versus Eigen on global accuracy | Eigen 2015 · SIFT Flow test | global accuracy: 86.8 -> 84.3 | Hurt |
| use our model instead of Long et al. on global accuracy | Long et al. · PASCAL Context val | global accuracy: 65.9 -> 62.4 | Hurt |
Closest papers by meaning
- Fully Convolutional Networks for Semantic Segmentation (TPAMI 2016 · similarity 0.91)
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (TPAMI 2017 · similarity 0.90)
- Exploring Cross-Image Pixel Contrast for Semantic Segmentation (ICCV 2021 · similarity 0.90)
- Revisiting Dilated Convolution: A Simple Approach for Weakly- and Semi-Supervised Semantic Segmentation (CVPR 2018 · similarity 0.90)
- Cross-Image Pixel Contrasting for Semantic Segmentation (TPAMI 2024 · similarity 0.90)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.89)
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.
