A Deep Quadruplet Network for Local Descriptor Learning

Xiaodong Wang,Yun Zhang,Dong Li,Kairui Chen,Xianxian Zeng,Yi Lyu

doi:10.1109/access.2019.2962624

Abstract

Generating a distinguishable feature for local patch is a main task in computer vision which aims at matching local patches. Recently, local patch descriptors from deep convolutional neural network (CNN) with a triplet loss have achieved promising performance. In this paper, we design a quadruplet loss, which can achieve a better result than other pairwise loss and triplet loss methods. Our loss is inspired by the thoughts of uniform distribution. It separates non-matching examples by using the hard sampled non-matching pairs in a batch, and simultaneously uses the random sampled non-matching examples to keep non-matching pairs to obey uniform distribution. A compact descriptor named QuadrupletNet is generated by combining the proposed quadruplet loss and L2Net CNN architecture. From our experiment, QuadrupletNet shows better performance on the Brown dataset and Hpatches dataset than Triplet loss methods on the same training set. The pre-trained QuadrupletNet is publicly available.

Highlights

In the last decade, how to generate a distinguishable feature for local patch is a main task in computer vision, because local patch matching is an important procedure for many computer vision problems, such as object recognition [2], 3Dreconstruction [3] and image stitching [4]
This paper proposes a novel quadruplet loss to train a compact descriptor named QuadrupletNet in the L2Net convolutional neural network (CNN) architecture
The histogram of the similarity of descriptors extracted by HardNet+ and HarNet(GOR)+ are shown as Figure 5(a) and Figure 5(b), and the result of DOAP+ and QuadrupletNet+ are shown in Figure 5(c) and Figure 5(d)

Summary

INTRODUCTION

How to generate a distinguishable feature for local patch is a main task in computer vision, because local patch matching is an important procedure for many computer vision problems, such as object recognition [2], 3Dreconstruction [3] and image stitching [4] Toward this end, so many handcraft feature descriptors are proposed, such as HOG [5], the famous SIFT [6] and SIFT’s variants [7]. In order to achieve better performance in learning based descriptors, [11] proposes L2Net, which can generate the output descriptor that can be matched in Euclidean space by L2 distance. A novel quadruplet loss for descriptor learning in cosine similarity space is proposed, which does not need the tunable parameter to balance different parts of the loss in different distance space. Different from the pair loss and triplet loss, the quadruplet loss does not need to decide the margin that separate the matching pairs and non-matching pairs

RELATED WORK

MODEL ARCHITECTURE AND TRAINING

EXPERIMENTAL EVALUATION

ETH DATASET

Findings

CONCLUSION