Research map / Architectures and efficiency / Neural architecture search
SasWOT: Real-Time Semantic Segmentation Architecture Search WithOut Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SasWOT, a training-free semantic segmentation architecture search framework that evolves a zero-cost proxy (combining multiple segmenter statistics with unary/binary operations) and uses it to search real-time segmenters without training candidates.
Problem
Semantic segmentation architecture search (SAS) is computationally expensive because existing methods train supernets via differentiable or RL search, and classification-designed zero-cost proxies correlate poorly with segmentation accuracy.
Approach
Evolve a customized zero-cost proxy for segmentation by building computation graphs over activations, gradients and weights with mathematical operators, using ranking correlation (Kendall's Tau) with segmentation accuracy as the fitness; then combine a searched proxy with NWOT via min-max normalization and run an evolutionary segmenter search with random initialized weights.
Key results
- On Cityscapes, SasWOT reaches 71.3% val mIoU and 69.8% test mIoU at 162.64 FPS, with a 1.8 GPU-hour search versus 48 GPU-hours for FasterSeg.
- On TransNAS-Bench-101 segmentation ranking, SasWOT attains a mean Kendall Tau of 64.08 in the micro space and 52.76 in the macro space, above NWOT (36.7 / 45.71).
- On CamVid (960x720), SasWOT achieves 64.3% mIoU versus 62.2% for NWOT and random search at similar FPS (~389).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine SasWOT- with the NWOT proxy using AVT normalization (ablation) | SasWOT segmenter search · Cityscapes test | mIoU (%): 66.7 -> 69.8 | Helped |
| Combine SasWOT- with the NWOT proxy using AVT normalization (ablation) | SasWOT segmenter search · CamVid test | mIoU (%): 58.7 -> 64.3 | Helped |
| Combine SasWOT- with the NWOT proxy using AVT normalization (ablation) | Proxy ranking on TransNAS-Bench-101 Semantic Segmentation · MICRO | Kendall Tau (%): 55.22 -> 64.08 | Helped |
| Combine SasWOT- with the NWOT proxy using AVT normalization (ablation) | Proxy ranking on TransNAS-Bench-101 Semantic Segmentation · MACRO | Kendall Tau (%): 28.92 -> 52.76 | Helped |
| Search segmenter architectures with SasWOT rather than NWOT | SasWOT · Cityscapes test | mIoU (%): 68.4 -> 69.8 | Helped |
| Search segmenter architectures with SasWOT rather than NWOT | SasWOT · CamVid test | mIoU (%): 62.2 -> 64.3 | Helped |
| Use evolutionary search instead of random search (ablation) | SasWOT proxy and architecture search · Search process | search convergence | Helped |
Surprising findings
- Classification-designed zero-cost proxies such as Fisher, Grad norm, and Grasp perform poorly on the segmentation ranking benchmark, with some Kendall Tau values below 10%.
Limitations
- Search space is inherited from FasterSeg and the proxy is validated mainly on a single benchmark family (TransNAS-Bench-101) and two driving datasets.
- The combined proxy relies on AVT min-max normalization estimated from randomly sampled architectures, a static policy that the authors acknowledge limits real-time scoring.
- Reported CamVid comparison is specific to that dataset and the authors note it may not generalize.
Experimental setup
Datasets: TransNAS-Bench-101 (semantic segmentation) Cityscapes CamVid
Benchmarks: Cityscapes CamVid TransNAS-Bench-101
Metrics: mIoU FPS Kendall Tau Spearman Pearson FLOPs search cost (GPU hours)
Baselines: Auto-DeepLab GAS CAS DF-Seg FasterSeg Random NWOT Synflow BiSeNet ICNet DFANet Fast-SCNN
Builds on: Zero-cost proxies (NWOT) Synflow Auto-Zero FasterSeg search space TransNAS-Bench-101
Closest papers by meaning
- FasterSeg: Searching for Faster Real-time Semantic Segmentation (ICLR 2020 · similarity 0.88)
- Customizable Architecture Search for Semantic Segmentation (CVPR 2019 · similarity 0.87)
- Graph-Guided Architecture Search for Real-Time Semantic Segmentation (CVPR 2020 · similarity 0.86)
- SCTNet: Single-Branch CNN with Transformer Semantic Information for Real-Time Segmentation (AAAI 2024 · similarity 0.86)
- RMP-SAM: Towards Real-Time Multi-Purpose Segment Anything (ICLR 2025 · similarity 0.85)
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation (NeurIPS 2025 · 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.