I completed my internship in the company this week and resigned from Qingdao Intelligent Manufacturing Data Co., Ltd. The main work tasks this week are code handover and deploying my previous facial recognition code to personal computers, and setting up the corresponding environment and testing, which has been completed.
The collection of face recognition public data sets was also carried out this week, and the results are as follows:
1. casia-facev5：
The CASIA-FaceV5 Asian face data set has 500 people, each person has 5 pictures, a total of 2500 pictures, and the picture size is 640*480. The format of the data is more in line with the direction I want to go now, but there are too few photos for each person, only 5. Link: https://pan.baidu.com/s/1ia6LdbXv7LOAr9XirntUsg, extraction code: BXW2.
2. CMUVASC & PIE Face dataset
The data set contains 40,000 photos from 68 people, including photos under 13 posture conditions, 43 lighting conditions and 4 expressions of each person. The official website is http://vasc.ri.cmu.edu /idb/html/face/index.html, but it failed. The resources and descriptions I collected from the Internet are not uniform. Although each picture is annotated with a name, the number of photos of each person is very small. The total number of photos is not 40,000, and the photos are black and white. Some of the photos may It is multi-target, so it cannot be used:
3
200 people, 7 pictures each, classified, grayscale, 80x80 pixels. The first picture is a standard image without change, the second and fifth pictures are images with large posture changes, and the third and fourth pictures are images with small posture changes. It works just fine as a training set (comparison library), but there is no test set.
Link: https://pan.baidu.com/s/1o9Y8v9rcdHn9763Z3HnVqA Extraction code: v7pt
4.CAS-PEAL
CAS-PEAL is a database of 99,450 face images of 1,040 volunteers completed by the Institute of Computing Technology of the Chinese Academy of Sciences in 2003. This database covers changes in features such as posture, expression, decoration, lighting, background, distance and time. It is a data set that is very in line with my expectations. To divide the training set and the test set, download link:
https://pan.baidu .com/s/1KZtOk-sfNkxnNQfNsTgTrQ extraction code: jbp6
5.CASIA-WebFace
The author of CASIA-WebFace wrote a crawler software to crawl the photos and names on IMDb and save them. The number of photos for each person is different.
6. AFLW
Annotation has a lot of dirty data. Although each person may have multiple photos, they cannot be distinguished by tag files.
7.GENKI4K
Smiley face data set, the file is damaged and the photo cannot be opened
8.The Japanese Female Facial Expression (JAFFE) Database
213 images, 10 people. Each person is a group, each group contains 7 expressions, and each expression has about 3 or 4 sample pictures. Photo naming can be used as annotation
Link: https://pan.baidu.com/s/1hrICsVq Password: klve
9.FDDB
Face Detection Data Set and Benchmark is a public database maintained by the Department of Computer Science at the University of Massachusetts, providing a standard face detection evaluation platform for researchers from all over the world. It is one of the most authoritative face detection evaluation platforms in the world, containing 2845 images and a total of 5171 faces as a test set. Download address: https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/
10.JAFFE
The database consists of 10 Japanese women who made various expressions according to instructions in an experimental environment. The entire database has a total of 213 images of 10 people, all of whom are women, and each person makes 7 expressions.
11. Person identification in TV series
The face photos come from two TV series "Buffy the Vampire Slayer" and "The Big Bang Theory". The tags are in mat format files and the photos are in videvens format files, which is more troublesome to use.
https://cvhci.anthropomatik.kit.edu/~baeuml/downloads/CVPR2013_PersonID_data_v1.1.tar.bz2
11. Personally produced data sets
The comparison library is as follows.
The test set is a face image obtained from a surveillance camera, as follows:
Summarize
Currently The Japanese Female Facial Expression (JAFFE) Database, CASIA-WebFace, CAS-PEAL , and personally produced data sets are available, among which the most expected one is CAS-PEAL