Real-Time Semantic Segmentation Network Based on Regional Self-Attention

Fig. 5. Visualization results of the Cityscapes dataset. (a) Original image; (b) real segmentation result; (c) segmentation result of our network; (d) segmentation result of the ERFNet

Download full size

Table 1. Comparison experiment of the zoom ratio
View table
Table 1. Comparison experiment of the zoom ratio
Serial number Network MIoU /% FPS /frame
1 (1,1,1,1) 71.9 10
2 (4,2,2,1) 71.7 109
3 (4,4,2,1) 71.7 120
4 (8,4,2,1) 71.6 133

Table 2. Results of the ablation experiments
View table
Table 2. Results of the ablation experiments
Network MIoU /% FPS /frame
Original network 68.4 158
RSA 71.7 120
RSA+LCIA 72.3 115
RSA+CBAM 72.5 80
RSA+SENet 72.1 102
RSA+LCIA + Booster 73.1 115

Table 3. Comparison experiment of the down-sampling method
View table
Table 3. Comparison experiment of the down-sampling method
Network MIoU /% FPS /frame
AvgPool 70.5 126
MaxPool 70.4 132
RSA 71.7 120

Table 4. Experimental results of different networks
View table
Table 4. Experimental results of different networks
Network Pretrain MIoU /% FPS /frame
ENet no 58.3 76
ESPNet no 60.3 112
ERFNet no 68.0 41.7
ICNet PSPNet50 69.5 30
DABNet no 70.1 104
DFANet^* Xception A 70.3 --
DFANet Xception A 71.3 100
Ours ResNet-18 72.1/71.8 115/126

Tools

Get Citation

Copy Citation Text

Hailong Bao, Min Wan, Zhongxiang Liu, Mian Qin, Haoyu Cui. Real-Time Semantic Segmentation Network Based on Regional Self-Attention[J]. Laser & Optoelectronics Progress, 2021, 58(8): 0810018

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites