GALLERY SET、PROBE SET

本文深入探讨了人脸识别系统中关键集合的概念:训练集、参考图像集(gallery set)与测试图像集(probeset)。通过引用相关论文,解释了这些集合在评估协议中的角色与作用,强调了训练过程与测试阶段的分离,以及系统参数的不可调整性。

摘要生成于 C知道 ,由 DeepSeek-R1 满血版支持, 前往体验 >

Deng Cai的Learning a Spatially Smooth Subspace for Face Recognition 4.1节也有,和朱师弟讨论此处 gallery set等同于训练集,Probe set等同于测试集; Graph Embedding and Extensions A General Framework for Dimensionality Reduction的4.1.1节的前一段最后意义与前相同
该论文的19页有如下描述:在FERET 评估协议中,算法设计者 需要区分三个不同的集合:训练集,参考图像集(或者
原型图像集,gallery set)、测试图像集合(Probe set),其中gallery  集和probe 集供测试时使用。“
练必须在测试开始之前完成”暗示训练是离线完成的,算法不能根据Gallery 集来调整系统参数。 

具体见 http://parnec.nuaa.edu.cn/xtan/paper/TanXY-thesis-final.pdf
该论文的19页有如下描述:在FERET 评估协议中,算法设计者需要区分三个不同的集合:训练集 参考图像集(或者
原型图像集,gallery set)、测试图像集合(Probe set=test set),其中gallery 集和probe 集供测试时使用。“训
练必须在测试开始之前完成”暗示训练是离线完成的,算法不能根据Gallery 集来调整系统参数。

通过和Ran He老师讨论已经彻底搞清楚了:

比如他的CVPR 2012论文,用PCA+NN。人脸识别是一个开集问题,人脸验证verification是一个闭集问题。训练样本可能是甲乙,测试样本可能是丙丁。通过训练样本来学习PCA的投影向量。gallery相当于新的训练集,在Probe上测试。在The CAS-PEAL large-scale Chinese face database and baseline evaluations (TSMCA 2008)P155也有Training set、 gallery set、Probe set定义

摘自:http://www.cppblog.com/guijie/archive/2008/09/16/61952.html

Surveillance Face Recognition Challenge # ======================== Dataset Structure ========================== # "Training_Set": Ordered by face identities, i.e. each directory contains face images from a specific training identity. There are a total of 5,319 directories each is named by the corresponding identity. The image file name is in format of [PersonID]_[CameraID]_[ImageName].jpg. "Face_Identification_Test_Set" -"gallery": containing 60,294 gallery images from 5,319 test identities (IDs). -"mated_probe": containing 60,423 probe images from 5,319 test IDs, with mated gallery true match images in the "gallery" folder. -"unmated_probe": containing 12,1736 distractor probe face images without mated gallery true match images in the "gallery" folder, i.e. the open-set face identification scenario. -"gallery_img_ID_pairs.mat": the gallery image names and the corresponding face IDs -"mated_probe_img_ID_pairs.mat": the mated probe image names and the corresponding face IDs "Face_Verification_Test_Set": -"verification_images": containing 10,051 images, a subset of the whole test set, randomly sampled to build 5,320 positive and 5,320 negative pairs -"positive_pairs_names.mat": 5320_by_2 cell specifying the image pairs of 5,320 positive pairs -"negative_pairs_names.mat": 5320_by_2 cell specifying the image pairs of 5,320 negative pairs # ======================== Evaluation Instruction ========================== # *** Open-Set Face Identification Evaluation ("Face_Identification_Evaluation") *** 1. Extract features of "gallery" images according to the order defined by "gallery_img_ID_pairs.mat", put them in a matrix called "gallery_feature_map" ([image_number]_by_[feature_dimension]), and save the matrix in a mat file named "gallery.mat" 2. Extract features of "mated_pr
03-10
评论 1
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值