Project aims to perform an end to end visual relationship detection. We start by performing object detection on raw images of Open Images Dataset. Then we apply a neural symbolic approach that injects prior common knowledge into a MLP trained on detected bounding boxes.

The goal is to merge the approaches shown in the 2.5 VRD reference paper to improve the results.

Run utils.py in preproccesing folder to download the images from OID, remember to specify the percentage if you want to create a subset of the training set.

Run function create_data_subset.py to create a folder containing a subset of images according to the train images that are in images/images_train.