Unsupervised semantic Segmentation with Pose Prior (USP)

By: Max Midwinter
CVISS Labs, Dept. CEE, University of Waterloo
Rogers 5G Smart Infrastructure, 2021

Demo

Run main.py for sample program

Own Data

1. Call prepDefect ( ) in main.py
   prepDefect takes 4 parameters:

   • REF_DIR: dir containing the reference frame
   • SUB_DIR: dir containing other images of the defect
   • AUG_DIR: dir where the preprocessed images are saved
   • scale: % to resize the raw images (Depends on your GPU and number of filters)
     • Keep resolution below 500x500

2. Call scribbsDefect ( ) in main.py
   scribbsDefect takes 2 parameters:

   • AUG_DIR: dir where images are saved
   • scribbs: scribbles to feed in
     • usually leave at None
   • save: save output image
     • For Debug save image (saved in current directory)

Docker Serving

Install Docker

pip install docker

Pull Compatible Tensorflow Docker Image

# Let me know if this does not work...
docker pull tensorflow/tensorflow:2.4.2-gpu

Build Docker Image

Build docker image usp with your choice of tag

docker docker build -t usp:TAG .

Run Docker Image

docker run usp:TAG

If you are running on local computer this command will start a dev server. (i.e. http://172.17.0.2:5000/)

You can now push your docker image to a container registry of your choice and deploy a kubernetes service...

To take advantage of parallel inference... Run main_docker.py (with your API)