"Policemen vs Civilians" is an image classification model. We used Convolutional Neural Networks (CNN) to learn a binary classifier on a small dataset (<400 samples) of both Serbian policemen and civilians. This repo is written in Tensorflow.
To train the model, we used VScode and we executed code/train.py.
Note: If training more than one model, you should change the paths in lines 22 & 23 of code/train.py.
After training a model, to look at the loss and metric curves, open an Anaconda terminal and type:
cd path/to/policemen_vs_civilians
tensorboard --logdir tb_logs
In your browser, go to localhost:6006 to see the tensorboard UI.
To evaluate the performance of the model on the validation data, start by changing the model path to the desired model to evaluate in line 9 of code/evaluate.py.
Then, execute code/evaluate.py. The code will plot the model probabilities on the validation data. Based on that figure, decide on a threshold to use for Accuracy, Precision and Recall metrics.
The model can be used on completely new images (even user provided).
The images to be tested on need to be square shaped. In our experiments, we cropped images to make them squared, with focus on the upper body of the person in the image.
All images that you want to test on, need to be placed inside the folder testpics.
To start, change the model path to the desired model in line 9 of code/inference.py. The code will ask for a model threshold to use to make the classification decision, and it will ask for images to be tested one by one.
Positives are policemen (label = 1) and negatives are civilians (label = 0).
