/SOLOv2

SOLOv2: Dynamic, Faster and Stronger. achieve 37.2mAP on coco val (res50,36 epoch) Adding keypoints

Primary LanguagePythonApache License 2.0Apache-2.0

SOLOv2

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

Install

Please check SOLOv1 for installation instructions.

Training

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

Eval

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

Visual

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

Weights

链接：https://pan.baidu.com/s/1tj-E63y5P__nzVFoAKAk4w 提取码：14r4

Results

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).

Visualization（1 epoch)