Our semantic segmentation implementation is based on MMSegmentation and PVT segmentation. Thank the authors for their wonderful works.
Please note that we just simply follow the hyper-parameters of PVT which may not be the optimal ones for Context Cluster. Feel free to tune the hyper-parameters to get better performance.
Install MMSegmentation v0.19.0.
Prepare ADE20K according to the guidelines in MMSegmentation.
| Method | Backbone | mIoU | Download |
|---|---|---|---|
| Semantic FPN | CoC-small-4 | 36.6 | model |
| Semantic FPN | CoC-small-25 | 36.4 | model |
| Semantic FPN | CoC-small-49 | 36.3 | model |
| Semantic FPN | CoC-medium-4 | 40.2 | model |
| Semantic FPN | CoC-medium-25 | 40.6 | model |
| Semantic FPN | CoC-medium-49 | 40.8 | model |
To evaluate Context Cluster + Semantic FPN on a single node with 4 GPUs run:
dist_test.sh configs/sem_fpn/{configure-file} /path/to/checkpoint_file 4 --out results.pkl --eval mIoU
To train Context Cluster + Semantic FPN on a single node with 8 GPUs run:
dist_train.sh configs/sem_fpn/{configure-file} 8