Overview of unconstrained low resolution face recognition I: data sets for low resolution face recognition

      at present , Very low resolution face recognition in surveillance scenes is a very small research field , The available data sets are very limited . among SCface、Point and Shot、IJB-S、UCCSface、QMUL Survface and QMUL Tinyface It is an available benchmark data set for unconstrained very low resolution face recognition .

Heterogeneous face recognition benchmark data set （ Both contain high-resolution and local low resolution images ）：SCface、Point and Shot、UCCSface and IJB-S;

Homogeneous face recognition dataset （ A lot of images and identities ）：QMUL-Survface and QMUL-TinyFace.

 PS：Source： Data acquisition scenario ;

         Qquality： Types of images available ;（HR—— High resolution image ,LR—— Very low resolution image ,blur—— Blurred low resolution image ;

         Static image/video： Indicates whether the dataset has a still image , Whether it also contains video data ;

Dataset Links ：

1、 Heterogeneous face data sets

1.1 SCface

      SCface The image was taken in an uncontrolled indoor environment using five video surveillance cameras of different quality . The dataset contains 130 Of subjects 4160 A still image .

https://www.scface.org/

1.2 Point and Shoot

       Point and Shoot Including still images and videos . Still photos have 9376 Zhang , common 293 people . The data set also includes 2802 paragraph 265 Human simple action video , this 265 Human still image 293 A subset of people .

https://www.nist.gov/programs-projects/point-and-shoot-face-recognition-challenge-pasc

 1.3 UCCSface

      UCCS It shows the characteristics of operable face recognition scenes . The dataset contains data from 308 Individual 6337 Zhang image . Use the camera to get images . The camera is placed in the office , Focus on leaving the office 100m On the outdoor sidewalk , produce 18m Pixel scene image . Image with 100ms Shooting at intervals , It produces about 10 Photos with different focus , There are multiple views and expressions at each specific interval .

https://vast.uccs.edu/Opensetface/

1.4 IJB-S

       IJB-S Data set containing 202 Static images and videos of identities . This data set was collected with 350 A surveillance video ,5656 Registered images and 202 Registration videos .

 Face Challenges | NIST

2、 Homogeneous face dataset

2.1 QMUL-Survface

         QMUL-Survface It is built by developing low resolution face recognition . The dataset contains 463507 Face images ,15573 A different identity . Because there are any large number of non target people in open space and unlimited time , Therefore, face recognition is usually more difficult in open settings . In addition to the low resolution problem , There are other uncontrolled covariates and noises in this data set , For example, lighting changes 、 expression 、 Occlusion 、 Background clutter and compression artifacts . All these factors will lead to the uncertainty of reasoning to varying degrees .

QMUL-SurvFace

2.2 QMUL-TinyFace

       QMUL-TinyFace The dataset has 169403 Zhang Yuanyuan's low resolution face image （ Average 20×16 Pixels ）, come from 5139 A marked identity , be used for 1:N Identification test .TinyFace All low resolution faces in are collected from public network data , These data are in posture 、 light 、 It is collected under the condition that the occlusion and background are not controlled .

  summary

      There are two cases of face recognition with very low resolution ： Homogeneity and heterogeneity . In homogeneous face recognition , We match the image source domain from the same human face . At this time, the recognized image and other reference images are from the unconstrained low resolution face domain . In heterogeneous face recognition , We match images from different fields ： Low resolution image for recognition and reference high resolution library image . therefore , From the surveillance camera VLR There is a gap between the probe image and the high-resolution reference gallery image taken in a controlled environment . The resolution of the low resolution image used for recognition is 32*32 Or lower , The resolution of the high-resolution image of the reference library is 100*100 Or higher . However, there are domain differences between the recognized image and the reference resolution image under different conditions , Heterogeneous face recognition is the most difficult problem .