Unofficial implementation for EmbedMask instance segmentation
office: https://github.com/yinghdb/EmbedMask
arxiv: https://arxiv.org/abs/1912.01954
| config | bbox | mask | weight |
|---|---|---|---|
| MS_R_50_2x.yaml | 40.399 | 34.105 | google drive |
The performance is not very stable at present.
MS_R_50_2x.yaml, box AP 34% and seg AP 28% were reached after a brief training.
Under the same number of iterations, MS_X_101_3x.yaml gets box AP 37%, while seg AP is only 24%.
Here I use the CPU implementation, so the FPS is low.
The code is based on detectron2. Please check Install.md for installation instructions.
Follows the same way as detectron2.
Single GPU:
python train_net.py --config-file configs/EmbedMask/MS_R_101_3x.yaml
Multi GPU(for example 8):
python train_net.py --num-gpus 8 --config-file configs/EmbedMask/MS_R_101_3x.yaml
Please adjust the IMS_PER_BATCH in the config file according to the GPU memory.
Single GPU:
python train_net.py --config-file configs/EmbedMask/MS_R_101_3x.yaml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file
Multi GPU(for example 8):
python train_net.py --num-gpus 8 --config-file configs/EmbedMask/MS_R_101_3x.yaml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file
I trained about 10 epochs, a V100 takes about 2 days.
