This repository has the code corresponding to our EndoCV 2021 submission, which is associated to the paper:
Multi-Center Polyp Segmentation withDouble Encoder-Decoder Networks
Adrian Galdran, Gustavo Carneiro, Miguel A. González Ballester
EndoCV 2021 Workshop on Computer Vision in Endoscopy 2021
that you can find here.
The interesting bit here is that we force a neural network to sample data from a multi-site dataset in such a way that in a batch all centers are as represented as possible. If you are interested in doing something similar on your own work, please check out the prepare_endo_data.py file, where we preprocess the data so that we oversample minority centers in our training set, and then the dataloader in utils/get_loaders.py, which provides batches with centers on it, allowing us to monitor per-center performance during training (see train_centers.py for this).