Equivariant neural networks for image segmentation

Deep neural networks have by today been established as the goto candidate for semantic or instance segmentation at many scales and image modalities. The pressing challenge in supervised segmentation approaches remains to be the requirement of large annotated image datasets for good performance.
In recent years the expressive capabilities of neural networks have been demonstrated to improve through group convolutional operations which exploit existing symmetries present in the data.
The increased capacity for weight-sharing alongside gains in sample efficiency for training a neural network have led to the empirical success of equivariant neural networks. In our study, we propose and experiment on an equivariant U-net-based model for the task of image segmentation. In this talk, we will discuss our preliminary results on a synthetic datasets consisting of polygonal objects. The results indicate that the performance of our implementation of an equivariant network improves well beyond a vanilla Unet when exposed to symmetrical objects in data different scenarios.

References:

1. Taco S. Cohen, Max Welling, “Group Equivariant convolution networks”, arXiv preprint arXiv: 1602.07576, 2016.
2. Maurice Weiler and Gabriele Cesa, ”General E(2)-Equivariant Steerable CNNs”, NeurIPS 2019.