The data set used is LFW (LabeledFaces in the Wild Home), which and FDDB (Face Detection Data Set and Benchmark) are the two most authoritative data sets for face recognition, respectively for the two most basic problems in face recognition: Detection and identification , gives detailed test requirements and scoring criteria.
LFW (Face Alignment Dataset)
Unconstrained natural scene face recognition data set, the data set consists of more than 13,000 face pictures of famous people around the world with different orientations, expressions and lighting environments in Internet natural scenes, a total of 5,749 people , of which 1,680 people have 2 or 2 pictures Picture of the face above. Each face picture is distinguished by its unique name ID and serial number.
The LFW data set mainly tests the accuracy of face recognition. The database randomly selects 6000 pairs of faces to form face recognition picture pairs, of which 3000 pairs belong to the same person with 2 face photos, and 3000 pairs belong to different people. 1 face photo. During the testing process, LFW gives a pair of photos and asks the system under test whether the two photos are of the same person, and the system gives a "yes" or "no" answer. The face recognition accuracy can be obtained by the ratio of the system answer to the real answer of 6000 pairs of face test results.
FDDB (Face Detection Dataset)
Unconstrained Natural Scene Face Detection Dataset, which contains 5171 faces in 2845 images captured from faces in various natural scenes. Each face has its specified coordinate position.
The FDDB dataset mainly tests the accuracy of face detection. The face recognition algorithm needs to perform face detection on each image in the dataset, and perform position calibration on the detected faces. Then, according to the correct answer given by the data set itself, the number of correctly detected faces and the number of falsely detected faces are calculated to judge the quality of the face detection algorithm.
FDDB and LFW Test Instructions
1. For companies and research institutions whose face detection rate of FDDB can exceed 90%, and that of LFW’s face recognition accuracy rate can exceed 99%, it only shows that they have a certain basis for face recognition algorithms, but they cannot reflect the algorithm at all. true level;
2. With the further combination of deep learning technology and face recognition technology, the results of the database test can be fully verified through the corresponding data targeted learning and supercomputing cluster hardware stacking and repeated verification to achieve full marks. (LFW has been brushed to a full score of 99.77% in 2015)
3. Even if the recognition rate on the FDDB and LFW databases reaches full marks, it does not prove that the face recognition technology can be used in practical application scenarios. The actual application scenarios and database data are vastly different, like It is the difference between talking on paper and fighting a war;
Differences between FDDB and LFW testing and practical applications
Example of FDDB and LFW pictures:
Example of police combat pictures:
FDDB and LFW brushing experience
1. Relying on super-large-scale Internet face data (the database image sources of FDDB and LFW come from the Internet);
2. Rely on ultra-high-performance supercomputing clusters and GPU clusters to train deep networks (reduce training time and experience errors of deep face recognition algorithms);
3. Rely on the superposition of deep learning models of different depths and complexity. (Use another type of depth algorithm to make up for the error correction of some depth algorithm's erroneous data).
4. In the case of known test data and standard answers, targeted learning and training; (repeatedly targeted optimization)
Finally, to emphasize again, LFW and FDDB are tests of the nature of question banks, and their main function is to test whether a system can achieve basic face recognition capabilities. In other words, if all face recognition systems are 3-year-olds, LFW is the intelligence level test used to test whether these children are good enough for kindergarten. The reason why it is said to be a question bank is because the 6000 groups of network samples, 6000 photos, are fixed. Any system can perform targeted optimization on these 6000 groups of samples, so as to achieve the effect of brushing high scores.
The FDDB dataset mainly tests the accuracy of face detection. The face recognition algorithm needs to perform face detection on each image in the dataset, and perform position calibration on the detected faces. Then, according to the correct answer given by the data set itself, the number of correctly detected faces and the number of falsely detected faces are calculated to judge the quality of the face detection algorithm.
