Apprentissage Contrastif et Segmentation : Exploitation des cartes de profondeur pour améliorer l’apprentissage de représentations

Ahmed Ben Saad, Axel Davy, Bastien Hell, Guillaume Erhard, Gabriele Facciolo

Abstract:

Unsupervised learning based on Contrastive Learning has received a lot of attention recently. This is due to the excellent results
obtained on a variety of subsequent tasks (in particular classification) on reference datasets (ImageNet, CIFAR-10, etc.) without the need
for a large amount of labeled samples. However, most reference contrastive learning algorithms (SimCLR, MoCo, BYOL) are not adapted to
downstream segmentation tasks. Here, we study a recently proposed variation of BYOL (called PixelCL), which is better adapted to segmentation and we propose an improvement to it by incorporating semantic information coming from the image depth in the pretraining. Our results show that PixelCL is able to use this new information, leading to better learned representations in an unsupervised manner.