Object detection for images of stage plays

Python

It is highly recommended to use conda to manage this project.
You will require two separate environments for the object detection and image colorization.

The object detection environment requirements can be installed using the .bachelor.txt file in the root of this repo: conda env create -n bachelor -f .bachelor.txt

The image colorization environment can be installed, by running:
conda env create -f deoldify/environment.yml

Custom libraries

Git submodule for tensorflow object detection models and DeOldify image colorizer.

Activate with:

git submodules update --init

To ensure the object detection and image colorization libraries work correctly, run:

# For object detection:
cd models/research/
protoc object_detection/protos/*.proto --python_out=.
cp object_detection/packages/tf2/setup.py .
python -m pip install -q .
cd ../../

# For image colorization
cd deoldify
python -m pip install -q .
cd ..

And then you should probably restart your shell (just to be safe).

Usage

Images

Images have to be in RGB colorspace, otherwise the notebook will not execute correctly.
If you want to use the original grayscale images, use the convert.sh script to make sure all your images are in RGB.

Structure

Folders:
input/grayscale - should contain all input images (jpg format)
input/colorized_stable - should contain all input images (jpg format) that have been colorized using the stable deoldify model input/colorized_artistic - should contain all input images (jpg format) that have been colorized using the artistic deoldify model results/ - will contain output images

Running

Make sure to always start the jupyter notebook server inside this folder (Object Detection). Otherwise the paths could be wrong.

Tips

No such file or directory: 'models\tmpXXXXXX' when trying to load colorizer model

Create a folder called 'models' in this directory and download the following two pretrained weights into that folder:
https://data.deepai.org/deoldify/ColorizeArtistic_gen.pth
https://www.dropbox.com/s/usf7uifrctqw9rl/ColorizeStable_gen.pth?dl=1

Font size of confidence score on resulting images

The font size of the displayed categories on detected objects can be changed:
Change font size (last answer)

Study results

Settings required to reproduce statistics sheet that works with the code given in the jupyter notebook.

Output Format: csv
All responses
Langauge: English
Range: Use default
Responses: Full answers