Introducing new #cvpr2020 work with S. Gidaris and team on a new self-supervised task: Learning Representations by Predicting Bags of Visual Words arxiv.org/abs/2002.12247 1/ 
@quobbe Inspired by NLP approaches, our method builds upon features from a self-supervised CNN (e.g. RotNet), which are used for computing a codebook of visual words and image-level Bag-of-Words (BoW) representations 2/
@quobbe Then, as a self-supervised task, we train another CNN to predict the BoW representation of an image given as input a perturbed version of that image. This forces the CNN to learn perturbation and context invariant features useful for downstream image-understanding tasks 3/
@quobbe We evaluate our method, BoWNet, on a range of datasets (CIFAR-100, MiniImageNet, ImageNet, Places205, VOC07 classif, VOC07+12 detection), outperforming ImageNet-trained supervised variants on VOC07+12 detection and getting similar performance on new classes, e.g. Places205 4/
@quobbe Most contrastive works focus on recognizing images under strong perturbation.
Differently from them, BowNet allows encoding multiple local visual concepts in the image. This opens the door for a wide set of tasks, e.g. urban images that differ only by few details. 5/
Differently from them, BowNet allows encoding multiple local visual concepts in the image. This opens the door for a wide set of tasks, e.g. urban images that differ only by few details. 5/
