Foveated Downsampling TechniquesDownload PDF

Anonymous

11 Sept 2019 (modified: 05 May 2023)Submitted to Real Neurons & Hidden Units @ NeurIPS 2019Readers: Everyone
TL;DR: We compare object recognition performance on images that are downsampled uniformly and with three different foveation schemes.
Keywords: foveation, fovea, neuroscience, neural, network, downsampling, saliency, perception, brain
Abstract: Foveation is an important part of human vision, and a number of deep networks have also used foveation. However, there have been few systematic comparisons between foveating and non-foveating deep networks, and between different variable-resolution downsampling methods. Here we define several such methods, and compare their performance on ImageNet recognition with a Densenet-121 network. The best variable-resolution method slightly outperforms uniform downsampling. Thus in our experiments, foveation does not substantially help or hinder object recognition in deep networks.
4 Replies

Loading