Bachelor Thesis - Counting Crates in Images Author: Petr Mičulek Date: 19 May 2021

The scripts provide 3 main use-cases

1. Generate cutouts datasets from source photos.

2. Train a CNN model

3. Evaluate a trained CNN model

The package requirements can be installed like:

pip install -r requirements.txt

###Generate cutouts datasets from source photos

To run the dataset generation, either:

a) Run generation of a single cutouts dataset (training + validation)

Uses the 128x sample size

python3 src_util/generate_dataset.py -f -b -c 128 -p 500 -s 25
python3 src_util/generate_dataset.py -b -c 128 -r -p 500 -s 25
python3 src_util/generate_dataset.py -f -b -c 128 -p 500 -s 25 -v
python3 src_util/generate_dataset.py -b -c 128 -r -p 500 -s 25 -v

b) Run generate_all_datasets.py

• Takes a few minutes, creates 18 dataset versions. Not necessary for a single training/evaluation run.

###Evaluate a trained CNN model

The final model is by default set for running evaluation.

The model weights and outputs folder names contain the model training run name: 64x_d1-3-5-7-9-11-1-1_2021-05-10-05-53-28_full

The code is best run by parts in an interactive environment:

ipython -i src/eval_run.py

###Train a CNN model

python3 src/training_run.py