Dividing the Neighbors is Not Enough: Adding Confusion Makes Local Descriptor Stronger

Abstract

Descriptors with learned features, other than descriptors with handcrafted features, have been popular in the research community. In this paper, we propose a novel piecewise loss function to improve the performance of learned descriptor, which is inspired by the nearest neighbor distance ratio matching strategy (the matching criterion for SIFT). In our strategy, if the distance ratio between positive and closest negative example is larger than a threshold, the loss will be chosen as HardNet loss; If the ratio is smaller than the threshold, the loss will be confusion. Applying the proposed loss function to the L2Net and HardNet architecture, the proposed descriptor is named ConfusionNet. Specifically, we find that adding the confusion to train local descriptors can achieve better performance. Comparing with previous works, we show in the experiments that ConfusionNet achieves state-of-the-art results in patch retrieval, patch verification, and image matching.

Keywords

• Convolutional neural network (CNN)
• learned descriptors
• confusion loss