SOLOv2-detectron2

Unofficial implementation for SOLOv2: Dynamic, Faster and Stronger instance segmentation.

Log

2020/6/12

At present, there are some bugs in the training code, leading to poor performance. Due to the lack of GPU, it is difficult to timely fix these bugs, so we should use it carefully.

2020/6/4

config	bbox	mask	weight
MS_R_50_2x.yaml	37.486	35.953	google drive

There are still a few bugs, "Person" is completely ignored, so performance should be higher than it is now.
Like this:

Now training code has been fixed, and the inference will not be affected.

Install

The code is based on detectron2. Please check Install.md for installation instructions.

Training

Follows the same way as detectron2.

Single GPU:

python train_net.py --config-file configs/MS_R_50_2x.yaml

Multi GPU(for example 8):

python train_net.py --num-gpus 8 --config-file configs/MS_R_50_2x.yaml

Please adjust the IMS_PER_BATCH in the config file according to the GPU memory.

Inference

First replace the original detectron2 installed postprocessing.py with the file.

Single GPU:

python train_net.py --config-file configs/MS_R_50_2x.yaml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file

Multi GPU(for example 8):

python train_net.py --num-gpus 8 --config-file configs/MS_R_50_2x.yaml --eval-only MODEL.WEIGHTS /path/to/checkpoint_file

Demo

cd demo/
python demo.py --config-file ../configs/MS_R_50_2x.yaml \
  --input /path/to/input_image \
  [--other-options]
  --opts MODEL.WEIGHTS /path/to/checkpoint_file

Results

2020/6/4

MS_R_50_2x.yaml