Reproducibility Challenge @ NeurIPS 2019: Unsupervised Object Segmentation by Redrawing

Adrian M. Chmielewski-Anders, Mats Steinweg, Bas T. Straathof

Abstract: Object segmentation has long been a topic of great interest in the field of computer vision. One of its greatest challenges is to overcome the need for costly, large, labeled data sets with pixel-level annotations. Chen et al., propose an adversarial model called ReDO, which is capable of unsupervised foreground-background segmentation in small images and can possibly be extended to multiple-class segmentation. The purpose of this report is to critically examine the reproducibility of the work by Chen et al., within the framework of the NeurIPS 2019 Reproducibility Challenge. The experiments described in this report partly corroborate the results of the original study. Moreover, the report offers a TensorFlow 2.0 implementation of ReDO and points out several flaws in the model description in the preprint of the original work.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=r1e95rSlUB