Pre-trained CNNs as Visual Feature Extractors: A Broad Evaluation

In this work, we perform a wide-ranging evaluation of Convolutional Neural Networks (CNNs) as feature extractors for matching visual features under large changes in appearance, perspective, and visual scale. Our evaluation covers 82 different layers from twelve different CNN architectures belonging to four families: AlexNets, VGG Nets, ResNets, and DenseNets. To our knowledge, this is the most comprehensive analysis of its kind in the literature. We find that the intermediate layers of DenseNets serve as the best feature extractors overall, providing the best overall trade-off of robustness to feature size. Moreover, we find that for each network, the later-intermediate layers provide the best performance, regardless of the total number of layers in the network.