The code is an unofficial pytorch implementation of [SOLOv2: Dynamic, Faster and Stronger] (https://arxiv.org/abs/2003.10152)
based on https://github.com/Epiphqny/SOLOv2
Please check SOLOv1 for installation instructions.
Follows the same way as SOLOv1.
single GPU:
python tools/train.py configs/solov2/solov2_r50_3x.py
multi GPU (for example 8):
./tools/dist_train.sh configs/solov2/solov2_r50_3x.py 8
Follows the same way as SOLOv1.
single GPU:
python tools/test_ins.py configs/solov2/solov2_r50_3x.py work_dirs/solo_r50_3x/latest.pth --show --out results_solo.pkl --eval segm
Follows the same way as SOLOv1.
single GPU:
python tools/test_ins_vis.py configs/solov2/solov2_r50_3x.py work_dirs/solo_r50_3x/latest.pth --show --save_dir work_dirs/vis_solo
链接：https://pan.baidu.com/s/1tj-E63y5P__nzVFoAKAk4w 提取码：14r4
After training 36 epochs(3x) on the coco dataset using the resnet-101 backbone, the mAP is 37.2 on COCO val-dev2017 dataset. In the original paper, the model achieves 38.8 after 72 epochs(6x).
