Research map / Detection and segmentation / CNN features
Scaling Semantic Segmentation Beyond 1K Classes on a Single GPU
Shipra Jain, Danda Pani Paudel, Martin Danelljan, Luc Van Gool
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 |
|---|---|---|---|
| sample nearest-neighbour negatives instead of random (ablation) | embedding segmentation · Cityscapes (MobileNet) | mIoU: 64.49 -> 67.20 | Improved |
| add normalization of class embeddings (ablation) | embedding segmentation · Cityscapes (MobileNet) | mIoU: 67.20 -> 72.56 | Improved |
| add max-margin regularization loss (ablation) | embedding segmentation · Cityscapes (MobileNet) | mIoU: 72.56 -> 73.03 | Improved |
| increase number of nearest neighbours k from 4 to 8 (ablation) | embedding segmentation · Cityscapes (MobileNet) | mIoU: 71.35 -> 71.08 | No clear effect |
| increase embedding dimension d from 4 to 7 (ablation) | embedding segmentation · Cityscapes | mIoU | Improved |
| replace d-channel embedding with C-channel output | DeepLabV3+ · COCO+LVIS (1284 classes) | mIoU: 1.68 -> 4.57 | Improved |
| replace d-channel embedding with C-channel output | DeepLabV3+ · COCO+LVIS (1284 classes) | FwIoU: 22.66 -> 39.67 | Improved |
| replace d-channel embedding with C-channel output | DeepLabV3+ ResNet50 · ADE20k | mIoU: 38.93 -> 38.29 | No clear effect |
| replace d-channel embedding with C-channel output | DeepLabV3+ MobileNet · Cityscapes | mIoU: 72.11 -> 73.03 | No clear effect |
| add gradient accumulation (ablation) | baseline · COCO+LVIS | mIoU: 1.68 -> 2.76 | Improved |
| add group normalization (ablation) | baseline · COCO+LVIS | mIoU: 1.68 -> 5.15 | Improved |
| add group normalization (ablation) | ours · COCO+LVIS | mIoU: 4.57 -> 6.26 | Improved |
Closest papers by meaning
- Large Kernel Matters -- Improve Semantic Segmentation by Global Convolutional Network (CVPR 2017 · similarity 0.89)
- Concept Mask: Large-Scale Segmentation from Semantic Concepts (ECCV 2018 · similarity 0.89)
- Lightweight and Progressively-Scalable Networks for Semantic Segmentation (IJCV 2023 · similarity 0.88)
- Boosting Semantic Segmentation from the Perspective of Explicit Class Embeddings (ICCV 2023 · similarity 0.88)
- DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs (TPAMI 2017 · similarity 0.88)
- Modeling the Background for Incremental and Weakly-Supervised Semantic Segmentation (TPAMI 2021 · 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.
