The repo contains code and data for my blog post
It provides useful tips on how not to mess up the Computer Vision model in production. I cover 4 real cases that I've met while working as an ML engineer. Each is related to image processing before passing it to the algorithm (mainly neural networks):
- Orientation stored in EXIF
- Non-standrad color profile
- Differences in image libraries
- Resizing algorithms
There is a Dockerfile with an environment to run the code.
One can either build and run it manually or use Makefile
make build
make runIt will start jupyter notebook which woule be available at localhost:8888.
After that feel free to open and run the notebook in notebooks/cv-prod-tips.ipynb