Here's a quick rundown of how I used the yolact (a well written semantic segmentation network) to train a detector and classifier for my face.
LabelMe Setup (in anaconda prompt)
conda create --name=labelme python=3.6
conda activate labelme
pip install labelme
git clone https://github.com/wkentaro/labelme
pip install pycocotools
LabelMe Usage
labelme
(label your images into separate folders: train, validate, test)
(create a label.txt each folder with classes)
Convert Labels to COCO format (in anaconda prompt)
.\labelme2coco.py <path to folder with pics>\nick_train\ <output folder>\nick_train_coco --labels <path to class list>\label.txt
(repeat for each folder)
(copy output folders into yolact/data/)
Configure dataset in yolact
configure yoloact/data/config.py with your dataset according to either of these:
https://github.com/dbolya/yolact
https://www.immersivelimit.com/tutorials/train-yolact-with-a-custom-coco-dataset
Train custom network
./train.py --config=yolact_resnet50_nick_dataset_config
If you get a cpu missing error, see the post below about fixing CPU/GPU error then try again.
Run the custom network
(webcam)
python eval.py --trained_model=./weights/yolact_plus_resnet50_nick_dataset_535_1070_interrupt.pth --config=yolact_resnet50_nick_dataset_config --score_threshold=0.15 --top_k=15 --video_multiframe=4 --video=0
(files)
python eval.py --trained_model=./weights/yolact_plus_resnet50_nick_dataset_535_1070_interrupt.pth --config=yolact_resnet50_nick_dataset_config --score_threshold=0.3 --top_k=15 --images=./data/nick_test:output_images
-----------------------------------------
References:
https://github.com/dbolya/yolact
https://github.com/wkentaro/labelme
The Papers
https://arxiv.org/pdf/1904.02689.pdf
https://arxiv.org/pdf/1912.06218.pdf
For Making Datasets
https://github.com/dbolya/yolact/issues/70#issuecomment-504283008
Tutorial for custom training
https://www.immersivelimit.com/tutorials/train-yolact-with-a-custom-coco-dataset
https://daddynkidsmakers.blogspot.com/2020/05/yolact-based-object-segmentation-with.html
Fixing training CPU/GPU error
https://github.com/dbolya/yolact/issues/664#issuecomment-878241658
