Hanawh / CondInst_mmdetection

Unofficial mmdetection implementation of CondInst.

Notes

CondInst source code : https://github.com/aim-uofa/AdelaiDet
SOLOv2 source code : https://github.com/WXinlong/SOLO
The model with ResNet-50 backbone(1x) achieves 28.7 mAP on COCO val2017 set.

• 4 gpu due to resource constraints
• NO FCOS bbox branch and centerness (Maybe because of this mAP drops)
• NO MS
• NO rel.coord
• NO auxiliary semantic segmentation
• Using matrix nms proposed by SOLOv2

If you need model, please contact me.

Environments (myself)

• python 3.7
• pytorch 1.1.0
• mmdetection v1.2.0 (please see the LINK for installation)
• mmcv 0.3.1

Train

# 4 gpu
./tools/dist_train.sh configs/condinst/condinst_r50_caffe_fpn_gn_1x_4gpu.py  4

# single gpu
python tools/train.py configs/condinst/condinst_r50_caffe_fpn_gn_1x_4gpu.py

Eval

python tools/test.py configs/condinst/condinst_r50_caffe_fpn_gn_1x_4gpu.py YOUR_CHECKPOINT_DIR  --out condinst.pkl --eval segm

Visualization

python tools/detect.py

some results you can see in 'outdata/'