Regressing Image Sub-Population Distributions with Deep Learning

Regressing the distribution of different sub-populations from a batch of images with learning algorithms is not a trivial task, as models tend to make errors that are unequally distributed across the different sub-populations. Obviously, the baseline is forming a histogram from the batch after having characterized each image independently. However, we show that this approach can be strongly improved by making the model aware of the ultimate task thanks to a density loss for both sub-populations related to classes (on three public datasets of image classification) and sub-populations related to size (on two public datasets of object detection in image). For example, class distribution was improved two-fold on the EUROSAT dataset and size distribution was improved by 10% on the PASCAL VOC dataset with both RESNET and VGG backbones. The code is released in the GitHub archive at achanhon/AdversarialModel/tree/master/proportion.