Optics and Precision Engineering, Volume. 32, Issue 8, 1212(2024)
Real-time urban street view semantic segmentation based on cross-layer aggregation network
Figures & Tables(12)
Fig. 1. Overall structure of CLANet （Cross-Layer Aggregation Network， CLANet）
Table 1. Ablation study of CLA-PPM on the Cityscapes validation
View tableView in ArticleTable 1. Ablation study of CLA-PPM on the Cityscapes validation
Baseline DAPPM CLA-PPM FLOPs/G Params/M mIoU/% Speed/FPS Sparsity=2 Sparsity=3 √ 97.7 14.2 74.5 96 √ √ 98.8 15.53 75.2 83 √ √ 98.0 14.68 75.0 89 √ √ 97.9 14.52 74.9 90 √ √ √ 98.2 15.0 75.3 85
Table 2. Ablation study of CLA-Net on the Cityscapes validation
View tableView in ArticleTable 2. Ablation study of CLA-Net on the Cityscapes validation
Baseline CLAM CLA-PPM MSFM FLOPs/G Params/M mIoU/% Speed /FPS √ 97.7 14.2 74.5 96 √ √ 103.1 14.5 75.2 86 √ √ 98.2 15.0 75.3 85 √ √ 98.3 14.22 74.8 93 √ √ √ 103.6 15.32 75.7 77 √ √ √ 98.8 15.04 75.5 83 √ √ √ 103.7 14.53 75.4 84 √ √ √ √ 104.3 15.35 76.0 75
Table 3. Comparison of accuracy and speed of Cityscapes
View tableView in ArticleTable 3. Comparison of accuracy and speed of Cityscapes
Method Reference Resolution mIoU/% #FPS
（PyTorch）
#FPS
（TensorRT）
Val Test GAS［ 19 ］CVPR2020 769×1 537 — 71.8 108.4 — HMSeg［ 20 ］BMVC2020 768×1 536 — 74.3 83.2 — DCNet［ 21 ］ICPR2021 512×1 024 — 71.2 142 — HyperSeg-M［ 22 ］CVPR2021 1 024×2 048 76.2 75.8 36.9 — RELAXNet［ 23 ］Neurocomputing2022 512×1 024 — 74.8 64 — FPANet C［ 24 ］APPL INTELL2022 1 024×2 048 — 75.9 31 — BiAttnNet［ 25 ］SPL2022 512 × 1 024 — 74.7 89.2 — LETNet［ 26 ］T-ITS2023 512×1 024 — 72.8 150 — SRDENet［ 27 ］IET COMPUT VIS2023 512×1 024 — 75.4 65 — BiSeNetV2［ 28 ］IJCV2021 512×1 024 73.4 72.6 — 156 BiSeNetV2-L［ 28 ］IJCV2021 512×1 024 75.8 75.3 — 47.3 FasterSeg［ 29 ］arXiv2019 1 024×2 048 73.1 71.5 — 163.9 STDC1-Seg50［ 9 ］CVPR2021 512×1 024 72.2 71.9 — 250.4 STDC1-Seg75［ 9 ］CVPR2021 768×1 536 74.5 75.3 — 126.7 CPANet-T50［ 30 ］CAC 2022 512 × 1 024 — 72.5 — 234.5 BiSeNetV3-50［ 31 ］Neurocomputing2023 512×1 024 73.4 73.5 — 244.3 CLANet-50（Ours） — 512×1 024 73.3 73.0 143 294 CLANet-75（Ours） — 768×1 536 76.0 75.8 75 164
Table 4. Comparison of accuracy and speed of CamVid
View tableView in ArticleTable 4. Comparison of accuracy and speed of CamVid
Method Reference Resolution mIoU/% #FPS
（PyTorch）
#FPS
（TensorRT）
CAS［ 32 ］CVPR2019 720×960 71.8 169 — GAS［ 19 ］CVPR2020 720×960 72.8 153.1 — DSANet［ 33 ］ExpertSyst. Appl. 2021 720×960 69.9 75.3 — FSFNet［ 34 ］IEEE T INSTRUM MEAS2021 720×960 75.1 91 — RELAXNet［ 23 ］Neurocomputing2022 720×960 71.2 79 — FPANet B［ 24 ］APPL INTELL2022 720×960 72.9 88 — LETNet［ 26 ］T-ITS2023 720×960 70.5 200 — SRDENet［ 27 ］IET COMPUT VIS2023 720×960 74.8 78.3 — BiSeNetV2［ 28 ］IJCV2021 720×960 72.4 — 124.5 BiSeNetV2-L［ 28 ］IJCV2021 720×960 73.2 — 32.7 STDC1-Seg50［ 9 ］CVPR2021 720×960 73.0 — 197.6 CPANet-T［ 30 ］CAC 2022 720×960 73.9 — 213.9 BiSeNetV3［ 31 ］Neurocomputing2023 720×960 75.1 — 198.4 CLANet（Ours） — 720×960 74.8 116 239
Tools
Get Citation
Copy Citation Text
Zhiqiang HOU, Minjie CHENG, Sugang MA, Minjie QU, Xiaobao YANG. Real-time urban street view semantic segmentation based on cross-layer aggregation network[J]. Optics and Precision Engineering, 2024, 32(8): 1212
Paper Information
Category:
Received: Oct. 21, 2023
Accepted: --
Published Online: May. 29, 2024
The Author Email: CHENG Minjie (rebu1999@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20