pandinosaurus/semanticsegmentation

A framework for training segmentation models in pytorch on labelme annotations with pretrained examples of skin, cat, and pizza topping segmentation

Semantic Segmentation

Overview

This project started as a replacement to the Skin Detection project that used traditional computer vision techniques. This project implements two models,

• FCNResNet101 from torchvision for accurate segmentation
• BiSeNetV2 for real-time segmentation

These models are trained with masks from labelme annotations. As labelme annotations allow for multiple categories per a pixel we use multi-label semantic segmentation. Both the accurate and real-time models are in the pretrained directory.

Getting Started

The pretrained models are stored in the repo with git-lfs, when you clone make sure you've pulled the files by calling,

git lfs pull

or by downloading them from github directly. This project uses conda to manage its enviroment; once conda is installed we create the enviroment and activate it,

conda env create -f enviroment.yml
conda activate semantic_segmentation

. On windows; powershell needs to be initialised and the execution policy needs to be modified.

conda init powershell
Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser

Pre-Trained Segmentation Projects

This project comes bundled with several pretrained models, which can be found in the pretrained directory. To infer segmentation masks on your images run evaluate_images.

# to display the output
python evaluate_images.py --images ~/Pictures/ --model pretrained/model_segmentation_skin_30.pth --model-type FCNResNet101 --display
# to save the output
python evaluate_images.py --images ~/Pictures/ --model pretrained/model_segmentation_skin_30.pth --model-type FCNResNet101 --save

To run the real-time models change the --model-type,

# to display the output
python evaluate_images.py --images ~/Pictures/ --model pretrained/model_segmentation_realtime_skin_30.pth --model-type BiSeNetV2 --display
# to save the output
python evaluate_images.py --images ~/Pictures/ --model pretrained/model_segmentation_realtime_skin_30.pth --model-type BiSeNetV2 --save

Skin Segmentation

This model was trained with a custom dataset of 150 images taken from COCO where skin segmentation annotations were added. This includes a wide variety of skin colours and lighting conditions making it more robust than the Skin Detection project. This model detects,

• skin

Pizza Topping Segmentation

This was trained with a custom dataset of 89 images taken from COCO where pizza topping annotations were added. There's very few images for each type of topping so this model performs very badly and needs quite a few more images to behave well!

• 'chilli', 'ham', 'jalapenos', 'mozzarella', 'mushrooms', 'olive', 'pepperoni', 'pineapple', 'salad', 'tomato'

Cat and Bird Segmentation

Annotated images of birds and cats were taken from COCO using the extract_from_coco script and then trained on.

• cat, birds

Training New Projects

To train a new project you can either create new labelme annotations on your images, to launch labelme run,

labelme

and start annotating your images! You'll need a couple of hundred. Alternatively if your category is already in COCO you can run the conversion tool to create labelme annotations from them.

python extract_from_coco.py --images ~/datasets/coco/val2017 --annotations ~/datasets/coco/annotations/instances_val2017.json --output ~/datasets/my_cat_images_val --categories cat

Once you've got a directory of labelme annotations you can check how the images will be shown to the model during training by running,

python check_dataset.py --dataset ~/datasets/my_cat_images_val
# to show our dataset with training augmentation
python check_dataset.py --dataset ~/datasets/my_cat_images_val --use-augmentation

. If your happy with the images and how they'll appear in training then train the model using,

python train.py --train ~/datasets/my_cat_images_train --val ~/datasets/my_cat_images_val --model-tag segmentation_cat --model-type FCNResNet101

. This may take some time depending on how many images you have. Tensorboard logs are available in the logs directory. To run your trained model on a directory of images run

# to display the output
python evaluate_images.py --images ~/Pictures/my_cat_imgs --model models/model_segmentation_cat_30.pth --model-type FCNResNet101 --display 
# to save the output
python evaluate_images.py --images ~/Pictures/my_cat_imgs --model models/model_segmentation_cat_30.pth --model-type FCNResNet101 --save