Reproducibility study - Counterfactual Generative Networks

Anonymous

Abstract:

In this work, we perfom a replication study of the paper Counterfactual Generative Networks by Axel Sauer, Andreas Geiger. We worked on the following claims;
(1)Counterfactual generative network (CGN) can generate high-quality counterfactual images with direct control over shape, texture, and background.
(2)Using generated counterfactual images in training data set improves the classifier’s out-of-domain robustness.
(3)Using generated counterfactual images in the training data set only marginally degrades overall accuracy.

Paper Url: https://openreview.net/forum?id=BXewfAYMmJw

Paper Venue: ICLR 2021