Summary
This paper was published on ICONIP21 , and you can download it from here.This paper mainly proposes a method that improves the robustness of features through clustering, and SOTA results are obtained on ReID data sets and Fashion Retrieval data sets.
这篇论文发表在ICONIP 21上,你可以在这里下载到这篇论文。这篇文章主要提出了通过聚类提高特征鲁棒性的方法,在ReID数据集和Fashion Retrieval数据集上取得了SOTA的结果。
Abstract
Image retrieval task consists of finding similar images to a query image from a set of gallery (database) images.A large portion of current research focuses on creating more robust features and modifying objective functions, usually based on Triplet Loss.In this paper we propose to use the mean centroid representation both during training and retrieval. Comprehensive experiments conducted on two ReID and Fashion Retrieval datasets demonstrate effectiveness of our method, which outperforms the current state-of-the-art.
图像检索是在gallery中找到那些和query相似图片的任务。目前大部分的研究关注:1)构建更具鲁棒性的特征,2)修改目标函数,通常基于Triplet Loss。在本文中,我们提出了一种在同时在训练阶段和检索阶段使用的Centroid(形心)度量。实验表明,我们的方法在两个ReID 和 Fashion Retrieval 数据集上取得了state-of-the-art结果。
Method
(a) Centroid-based retrieval
(b) Instance-based retrieval
The Figures above is the Comparison of centroid-based and instance-based retrieval.
传统的TripletLoss计算正负样本和query间的距离,通过学习参数使正样本和query间的距离缩短,负样本和query间的距离扩大。而作者则计算正负样本的形心和query间的距离。形心表征聚合了一类的表征,使用形心表征,1)能减少计算,2)对离群值和噪声标签有更高的鲁棒性,3)训练更快,4)比传统TripletLoss有更好的性能。
Centroid Trilet Loss
传统的Triplet Loss function 计算样本-样本间的距离,鲁棒性较差
L t r i p l e t = [ ∥ f ( A ) − f ( P ) ∥ 2 2 − ∥ f ( A ) − f ( N ) ∥ 2 2 + α ] + \mathcal L_{triplet} = [\|f(A)-f(P)\|^2_2-\|f(A)-f(N)\|^2_2+\alpha]_+ Ltriplet=[∥f(A)−f(P)∥22−∥f(A)−f(N)∥22+α]+