This is a Jittor implementation of CCNet for semantic segmentation.
You should install Jittor under the guidance of Jittor official.
To train with default batch-size 16 setting, you only need to specify the desired log directory, and run:
mpirun -np 4 python train.py --log_dir <log directory>
To test a model, you can specify the model checkpoint directory and the model's backbone, and run:
python test.py --ckpt_dir <log directory> --model_backbone <van or resnet>
To visualize a result, you can specify the picture's directory and the model, and run:
python visualize.py --ckpt_dir <log directory> --model_backbone <van or resnet> --pic_dir <picture directory> --save_dir <save directory>
| Backbone | Checkpoints |
|---|---|
| ResNet101-C | TsinghuaCloud |
| VAN-base | TsinghuaCloud |
| Settings | mIoU | Checkpoints |
|---|---|---|
| ResNet101-C, batch-size 16 | 41.72 (42.80 wi multiscale) | TsinghuaCloud |
| ResNet101-C, batch-size 8 | 38.85 | TsinghuaCloud |
| ResNet101-C, batch-size 8, dilated | 38.72 | TsinghuaCloud |
| ResNet101-C, batch-size 8, neighborhood | 36.77 | TsinghuaCloud |
| VAN-base, batch-size 16 | 36.82 | TsinghuaCloud |