Author Affiliations
College of Electrical Engineering and Control Science, Nanjing Tech University, Nanjing, Jiangsu 211816, Chinashow less
Fig. 1. Different convolution types. (a) Ordinary convolution; (b) dilated convolution; (c) dilated convolution after smoothing
Fig. 2. Dilated convolution and smoothing effect
[24]. (a) Dilated convolution; (b) dilated convolution after smoothing
Fig. 3. Structure of residual unit in convolution block
Fig. 4. Structural diagram of network with HC-LUM
Fig. 5. Comparison of segmentation accuracy between proposed segmentation method and other methods for 19 types of objects
Fig. 6. Influence of knowledge distillation method on accuracy of each segmentation network
Fig. 7. Comparison of segmentation accuracy for different network layers
Fig. 8. Segmentation results of proposed method. (a) Original images 1; (b) segmentation results of original images 1 (including enlarged images of partial detail); (c) original images 2; (d) segmentation results of original images 2
Method | MIOU /% | Frame rate /(frame·s-1) | Parameter /107 |
---|
ResNeXt-18+D-Cov | 72.3 | 36.3 | 1.25 | ResNeXt-18+DCSM | 73.7 | 36.1 | 1.39 | ResNeXt-18+DCSM+HC-LUM | 73.9 | 35.9 | 1.18 | Proposed | 76.8 | 40.2 | 1.18 |
|
Table 1. Corresponding results of four segmentation methods on Cityscapes dataset
Method | MIOU /% | Frame rate /(frame·s-1) |
---|
ResNeXt-18+D-Cov | 64.2 | 33.9 | ResNeXt-18+DCSM | 64.7 | 33.7 | ResNeXt-18+DCSM+HC-LUM | 65.1 | 32.5 | Proposed | 65.3 | 34.2 |
|
Table 2. Corresponding results of four segmentation methods on CamVid dataset
Method | MIOU /% | Frame rate /(frame·s-1) |
---|
ICNet | 69.5 | 30.3 | Two-column Net | 72.9 | 14.7 | LadderDenseNet | 72.82 | 31.0 | ESPNet | 60.3 | 112 | ERFNet | 68.0 | 11.2 | GUNet | 70.4 | 37.3 | Proposed | 76.8 | 40.2 |
|
Table 3. Comparison of proposed method with other segmentation networks (Cityscapes dataset)
Method | MIOU /% | Frame rate /(frame·s-1) |
---|
ICNet | 67.1 | 27.8 | PSPNet | 69.1 | 5.4 | Dilation10 | 65.3 | 4.4 | SegNet | 46.4 | 4.6 | ERFNet | 59.4 | 10.1 | GUNet | 61.8 | 31.3 | Proposed | 65.3 | 34.2 |
|
Table 4. Comparison of proposed method with other segmentation networks (CamVid dataset)