The code is an unofficial pytorch implementation of SOLO: Segmenting Objects by Locations
The code is based on mmdetection. Please check Install.md for installation instructions.
Follows the same way as mmdetection.
single GPU: python tools/train.py configs/solo/r50.py
multi GPU (for example 8): ./tools/dist_train.sh configs/solo/r50.py 8
The code only implements the simplest version of SOLO:
- without CoordConv
- using vanilla SOLO instead of Decoupled SOLO
- 3x training schedule
- using the default FPN featuremaps: in the paper it is with different specific strides and instance scale selection
- implemented the simplest mask-nms: as the authors did not describe it in detail in the paper, the implemented nms is slow, will improve it in the future.
- still in progress
After training 36 epochs on the coco dataset using the resnet-50 backbone, the mAP is 0.221 on COCO val2017 dataset:
Both good and bad results:
