Balanced sampling for an object detection problem - application to fetal anatomies detection

In this paper, we propose a novel approach to overcome the problem of imbalanced datasets for object detection tasks, when the distribution is not uniform over all classes. The general idea is to compute a probability vector, encoding the probability for each image to be fed to the network during the training phase. This probability vector is computed by solving some quadratic optimization problem and ensures that all classes are seen with similar frequency. We apply this method to a fetal anatomies detection problem, and conduct a thorough statistical analysis of the resulting performance to show that it performs significantly better than two baseline models: one with images sampled uniformly and one implementing classical oversampling.