Research map / Detection and segmentation / CNN features
Dense and Low-Rank Gaussian CRFs Using Deep Embeddings
Siddhartha Chandra, Nicolas Usunier, Iasonas Kokkinos
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 |
|---|---|---|---|
| make pairwise terms dense (fully connected) with low-rank embeddings, replacing sparse Potts G-CRF | Deeplab-v2 · PASCAL VOC 2012 test | mean IoU (%): 79.5 -> 80.3 | Improved |
| add dense Potts G-CRF to the base network | Deeplab-v2 · PASCAL VOC 2012 test | mean IoU (%): 79.0 -> 80.3 | Improved |
| add sparse G-CRF on top of dense Potts model | densepotts-G-CRF · PASCAL VOC 2012 test | mean IoU (%): 80.3 -> 80.4 | Improved |
| add dense Potts G-CRF in place of sparse Potts G-CRF | Deeplab-v2 · PASCAL Person Parts | mean IoU (%): 65.21 -> 66.27 | Improved |
| add sparse G-CRF on top of dense Potts model | densepotts-G-CRF · PASCAL Person Parts | mean IoU (%): 66.27 -> 66.31 | Improved |
| add dense Potts G-CRF without dense-CRF post-processing | Deeplab-V2 · PASCAL Person Parts | mean IoU (%): 64.94 -> 66.27 | Improved |
| add dense G-CRF to base network | Deeplab-v2 · HKU-IS | maximal F-measure: 0.916 -> 0.927 | Improved |
| add dense G-CRF to base network | Deeplab-v2 · PASCAL-S | maximal F-measure: 0.859 -> 0.872 | Improved |
| add sparse G-CRF on top of dense G-CRF | dense-G-CRF · PASCAL-S | maximal F-measure: 0.872 -> 0.864 | Hurt |
| deepen the pairwise stream from res2a to res4a (ablation) | dense-G-CRF (dim 128) · PASCAL VOC 2012 val | mean IoU (%): 76.81 -> 77.05 | Improved |
| deepen the pairwise stream from res2a to res4a (ablation) | densepotts-G-CRF (dim 1024) · PASCAL VOC 2012 val | mean IoU (%): 77.10 -> 77.21 | Improved |
| increase embedding dimension from 128 to 512 (ablation) | dense-G-CRF (res4a) · PASCAL VOC 2012 val | mean IoU (%): 77.05 -> 76.97 | No clear effect |
Closest papers by meaning
- Fast, Exact and Multi-scale Inference for Semantic Image Segmentation with Deep Gaussian CRFs (ECCV 2016 · similarity 0.93)
- Gaussian Conditional Random Field Network for Semantic Segmentation (CVPR 2016 · similarity 0.91)
- Regularized Frank-Wolfe for Dense CRFs: Generalizing Mean Field and Beyond (NeurIPS 2021 · similarity 0.88)
- Efficient Linear Programming for Dense CRFs (CVPR 2017 · similarity 0.88)
- Deep Spatio-Temporal Random Fields for Efficient Video Segmentation (CVPR 2018 · similarity 0.88)
- Higher Order Conditional Random Fields in Deep Neural Networks (ECCV 2016 · similarity 0.88)
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.
