Sihai network

Pengsi technology won the first place in three iccv2019 competitions, continuously deepening the res

The Biennial International Conference on computer vision from October 27 to November 2

Iccv2019 (International Conference on computer vision) was held in Seoul, South Korea. The International Conference on computer vision (iccv), the conference on computer vision pattern recognition (CVPR) and the European Conference on computer vision (ECCV) are also called the top three conferences on the direction of computer vision. For the first time, Pengsi technology won the first prize in three iccvlfr competitions and the special Creative Award issued by the organizing committee, which shows the strength of computer vision technology of Pengsi in landing application scenario.

Unconstrained static image face recognition and dynamic video face recognition is one of the most in-depth research topics in the field of computer vision, which has a wide range of application scenarios in video monitoring, biomedical and other fields. In recent years, new technologies and methods in various fields of computer vision have been put forward in many top-level computer vision conferences, and the face recognition technology based on deep learning has made phased progress.

Although many achievements have been made in the field of face recognition, the research and efforts of lightweight face recognition based on deep learning still need to be strengthened. In the face of the application of large database face recognition, it is still a challenge to find a lightweight and high-precision model that can perform well in the unlimited dynamic surveillance video scene.

Iccvlfr (lightweight face recognition challenge) is set up for this purpose. It has also become an important competition of iccv this year, attracting 292 teams from all over the world. Unlike other face recognition competitions, iccvlfr challenge strictly limits training data and test data. Under such strict conditions, it is very difficult to stand out from 292 teams.

Training and test data of iccvlfr competition

The iccvlfr challenge is divided into four competitions, each of which has its own limitations and focuses:

1. Protocol-1 (deepglint light) image face recognition lightweight recognition model, the computational complexity is less than 1gflops, the model size is less than 20MB, the data type is float32, and the feature dimension is 512 (FPR @ 1e-8);

2. Protocol-2 (deepglint large) image face recognition large-scale recognition model, the operation complexity is less than 30gflops, the data type is float32, and the feature dimension is 512 (FPR @ 1e-8);

3. Protocol-3 (iqiyi light) video face recognition lightweight recognition model, the computational complexity is less than 1gflops, the data type is float32, and the feature dimension is 512 (FPR @ 1e-4);

4. Protocol-4 (iqiyi large) video face recognition large-scale recognition model, the computational complexity is less than 30gflops, the data type is float32, and the feature dimension is 512 (FPR @ 1e-4).

Pengsi Singapore Research Institute team makes oral presentation

Finally, Shen shengmei team of Pengsi Singapore Research Institute put forward unsupervised learning method based on relation diagram to strengthen features, which ranked first in three competitions of iccvlfr challenge, i.e. lightweight image recognition, large-scale image recognition and lightweight video image recognition, leading research institutions and enterprises such as Microsoft Asia Research Institute, Automation Institute of Chinese Academy of Sciences, Jingdong, today's headlines, etc., and There is a big gap with them. In the light-weight image recognition competition, when the false alarm rate is one hundred million (1e-8), Pengsi technology has achieved 93.41%, nearly six percentage points higher than other competitors. In the lightweight video and image competition, 72.23% of Pengsi technology's results are nearly nine percentage points higher than those of other competitors.

Iccvlfr results

In the workshop of iccvlfr, the team of Singapore Research Institute of Pengsi made an oral report entitled 'agraphbasedunsurveisedfeatureclustering for facerecognition', elaborated the unsupervised learning method proposed by the team, and made the feature distribution of the same ID more compact and the distance between features of different IDS more scattered by using the relationship between the two test data, thus greatly The recognition accuracy is improved. The effectiveness of this method has been verified in ijb-c, YTF and CFP databases, and the accuracy of baseline model has been greatly improved.

Example flow chart of unsupervised learning method based on relation chart

In addition to the first place in the iccvlfr challenge, the performance of the latest algorithm model of Pengsi technology in the public face recognition data set ijb-c of NIST in the United States has also broken the world record. Ijb-c is the most scientific and comprehensive benchmark database in NIST face recognition open data set in the United States under unlimited conditions. Today, when the accuracy of LFW, cfp-fp and agedb-30 is generally saturated, ijb-c is the face recognition reference database closest to the actual scene of video surveillance.

Test results of Pengsi technology on nistijb-c data set

Not only in the field of face recognition, but also in recent years, Pengsi technology has made breakthroughs in such computer vision technologies as pedestrian re identification (Reid), video based re identification (video based Reid), real-time location and map building (SLAM), and won many world champions.

-In July, Pengsi technology set a new world record in the test of three major data sets of pedestrian recognition (Reid), market151, dukemtmc Reid and cuhk03.

-In August, Pengsi technology refreshed three major data sets of video based Reid: prid-2011, ilids-vid and Mars.

-In October, in the just concluded ismar2019, Pengsi technology won the third place in the vSLAM competition in the ar-slam challenge. Slam technology can be used in a variety of applications including autonomous driving, mobile robots, 3D reconstruction, augmented reality and hybrid reality. Vision slam is an important accumulation of Pengsi technology in computer vision technology, and will add more value to the company's current and future business. At the 2019 Shenzhen Security Expo, Pengsi Technology launched an unmanned patrol police car for security scenarios. The follow-up in-depth study of visualslam technology developed by Pengsi will gradually replace 3dlidar to complete the positioning and navigation of the unmanned patrol car, so as to reduce costs and expand the application field.

At present, Pengsi technology has set up research institutes in Beijing and Singapore, constantly bringing together global AI talents. It has fully self-developed and full stack computer vision technology, covering multiple research fields of computer vision. On this basis, Pengsi technology, on the one hand, is based on the existing business and business model, and combines with the company's development direction to carry out Scenario Oriented AI technology research and development and innovation; on the other hand, it constantly explores cutting-edge technology from a global perspective, so that the company always maintains the sensitivity and attention to breakthrough technology, and makes technical reserves for the company's efforts in the field of artificial intelligence and exploration of new business.