Research map / Detection and segmentation / CNN features
SegBlocks: Block-Based Dynamic Resolution Networks for Real-Time Segmentation
Thomas Verelst, Tinne Tuytelaars
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 |
|---|---|---|---|
| use average-sampling instead of zero-padding in BlockPad (ablation) | SegBlocks-RN18 (tau=0.4) · Cityscapes val | mIoU: 70.2% -> 76.3% | Improved |
| use average-sampling instead of strided subsampling for high-res neighbours in BlockPad (ablation) | SegBlocks-RN18 (tau=0.4) · Cityscapes val | mIoU: 75.6% -> 76.3% | Improved |
| use block size 128x128 instead of 256x256 (ablation) | SegBlocks-RN18 (tau=0.4) · Cityscapes val | mIoU: 75.2% -> 76.3% | Improved |
| use block size 64x64 instead of 128x128 (ablation) | SegBlocks-RN18 (tau=0.4) · Cityscapes val | mIoU: 76.3% -> 76.3% | No clear effect |
| enlarge policy network input from 512x256 (4 layers) to 1024x512 (5 layers) (ablation) | SegBlocks-RN18 policy network · Cityscapes val | mIoU: 76.3 -> 76.1 | No clear effect |
| shrink policy network to 128x64 input, 2 layers (ablation) | SegBlocks-RN18 policy network · Cityscapes val | mIoU: 76.3 -> 75.3 | Hurt |
| widen policy network features from 64 to 128 (ablation) | SegBlocks-RN18 policy network · Cityscapes val | mIoU: 76.3 -> 76.1 | No clear effect |
| add dynamic block resolution (tau=0.4) vs static baseline | SwiftNet-RN18 · Cityscapes val | GMACs: 104.1 -> 60.5 | Improved |
| add dynamic block resolution (tau=0.2) vs static baseline | SwiftNet-RN50 · Cityscapes (GTX 1080 Ti) | FPS: 16.3 -> 30.0 | Improved |
| dynamic resolution vs static lower input resolution of 1536x768 | SwiftNet-RN18 · Cityscapes val | mIoU: 1.7% lower mIoU for static baseline | Improved |
| add dynamic block resolution (tau=0.4) on CamVid | SwiftNet-EffL1 · CamVid test | mIoU: 75.1 -> 74.6 | Hurt |
| add dynamic block resolution (tau=0.2) on Mapillary Vistas | SwiftNet-EffL1 · Mapillary Vistas val | mIoU: 41.7 -> 39.8 mIoU, 20.7 -> 40.1 FPS | Mixed |
Closest papers by meaning
- FasterSeg: Searching for Faster Real-time Semantic Segmentation (ICLR 2020 · similarity 0.86)
- HyperSeg: Patch-Wise Hypernetwork for Real-Time Semantic Segmentation (CVPR 2021 · similarity 0.86)
- Dynamic Face Video Segmentation via Reinforcement Learning (CVPR 2020 · similarity 0.85)
- SegGPT: Towards Segmenting Everything in Context (ICCV 2023 · similarity 0.85)
- Lightweight and Progressively-Scalable Networks for Semantic Segmentation (IJCV 2023 · similarity 0.85)
- Reinforced active learning for image segmentation (ICLR 2020 · similarity 0.84)
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.
