A Graph-involved Lightweight Semantic Segmentation Network: GLNet.
Run cityscapes_train.py for training and cityscapes_eval.py for inference. The experiments were conducted on Ubintu OS with an RTX 4090 GPU, it might needs some modification for running the codes on Windows OS. Besides, the well-trained weight is stored as /checkpoint/cityscapes/modelbest.pth.
Results on Cityscapes val set
| Methods | Resolutions | mIoU(%) | Params |
|---|---|---|---|
| ESPNetv1 | 61.4 | 0.4 | |
| CGNet | 63.5 | 0.5 | |
| Lightset-Shuffle | 66.1 | 3.5 | |
| DGCNet(Res101) | 80.5 | 72.5 | |
| WaveMix- | 63.3 | 2.9 | |
| SegFormer(B0) | 62.6 | 7.7 | |
| PPL-LCNe | 66.0 | 3.0 | |
| TopFormer-Tiny | 66.1 | 1.4 | |
| MGD | 62.6 | 1.8 | |
| Ours | 66.8 | 1.5 |
Visualization on Cityscapes val set.
Xue Xia, Jiayu You, Yuming Fang. A Graph-involved Lightweight Semantic Segmentation Network. PRCV2023, Accepted.
This work was inspired by CGNet: A Light-weight Context Guided Network for Semantic Segmentation and Dual Graph Convolutional Network for Semantic Segmentation.