Reproducing CNN2: Viewpoint Generalization via a Binocular Vision

Manoosh Samiei, Rezvan Sherkati, Aarash Feizi

Abstract: In this project, we replicate the results of the paper “CNN2: Viewpoint Generalization via a Binocular Vision” for two datasets SmallNORB and ModelNet2D. We realize that, CNN2 gives results close to the original CNN model, referred to as “vanilla CNN” by the authors. We also perform an ablation study, by changing the network’s architecture, multiple setups, and hyperparameters in the proposed method. We observe that some modifications such as changing the CM pooling to max pooling, lower the accuracy as they have also claimed. However, there are some modifications that appear to be improving the accuracy, in contrast to what the authors claim in the paper.

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=rkl0BVrxLB