What next after the outlet Face recognition is?
For this problem, the industry seems to have reached a consensus. Since the AI face recognition capabilities beyond the human eyes academia and industry are turning to another more scientific and practical value of the subject - heavy pedestrian recognition (Person Re-identification, ReID).
Recently, according to map technology to achieve new breakthroughs in the field of ReID, technology refresh its proposed global industry's three leading industry data sets optimal performance (SOTA), algorithm performance so far reached the highest industry standards and greatly expanded algorithms and applications boundary.
NOTE: YITU algorithm result in time and space without using information that does not reflow re-optimization, i.e.: Re-ranking achieved under like conditions.
Remember the end of 2018 in accordance with plans to enter the smart voice, then setting a new record in Chinese recognition accuracy of speech recognition. In May 2019 launched the world's first cloud-visual AI chip, and "that is commercially released." In some technical fields, according to the industry standard map onto a new high point, and accelerate the industrialization of landing technology.
What are the key behind this?
Heavy pedestrian recognition (ReID), the face recognition "killer app."
In the actual scene transportation, industrial manufacturing and urban planning, 99% of the images are free of human faces - even if there's a face part is also extremely vague, only a few pixels in size, this time face recognition more limited role.
Heavy pedestrian recognition (ReID, also known as "re-pedestrian recognition"), refers to the pedestrian retrieved in multi-camera network equipment, the use of gait movement, physical characteristics such as more comprehensive information to identify the person, whether used alone or with the combination of face recognition, can play a greater value.
In addition to retail scenarios intelligent, intelligent transportation, smart cities and so often mentioned, application ReID technology will also make daily life more convenient: amusement park easier to find lost children, pets / household robot can accurately identify the owner or the back and provide customers with the corresponding service.
However, due to the need ReID images or videos taken from different cameras to identify with a character, and the scope covered by these cameras do not overlap each other, resulting in a lack of consistent information, and a different attitude screen characters, behavior and even appearance ( For example: the seating arrangement, sideways, back body) larger vary, at different times, the scene illumination, and the background covering different (there are often integrated background, character clothing other similar interference), but also a high resolution of the camera there are low, the position of the characters appear on the screen that are far forward, these are all ReID technology presents a great challenge.
ReID depth optimization algorithm framework, AutoML replace the manual tuning algorithm
Faced with this situation, according to the optimization algorithm framework ReID the depth chart, significantly improve the efficiency of the algorithm, by combining AutoML and other cutting-edge technologies to further innovation achieved automatically search and iterative model parameters, broke through the manual design-dependent algorithm and researcher tuning algorithms traditional development processes, while reducing labor costs, such that the stronger algorithm generalization performance.
Two key indicators of the three ReID according to data from research sets Market1501 map algorithm in the industry's most influential, DukeMTMC-ReID, on CUHK03, we will measure the performance of the algorithm "first hit rate" (Rank-1 Accuracy) and " all the mean average precision "(mean average precision, mAP) 6 item data upgrade, which fully demonstrates the technical strength in accordance with plans and further reinforces China AI lead position in this task.
It should be noted, first hit rate, only means that the algorithm can accurately identify the most easily recognizable or match goes on in a number of images, and can not reflect the true ability of the model, especially in response to the performance of complex scenes.
Therefore, the need to combine mAP value when evaluating ReID algorithm performance, which reflects the performance of the integrated retrieval system. MAP The higher the value, the better illustrate the utility of the system, not only can the whole Richard Richard accurate, better able to cope with multi-block, light, dark, blurring and so on.
AI research since the chip QuestCore help accelerate the commercialization of the world's leading ReID algorithm landing
In addition to algorithm performance, large-scale application of restrictions ReID landing another major reason for the commercialization of the existing terminal devices such as a camera operator is not strong enough force. It can be said, it has been the lack of force considered AI commercial landing pain points.
According to the chart in 2017 from the research cloud AI chip QuestCore (quest), and "that is commercially released" in May 2019. QuestCore is the world's first cloud-visual AI chip, providing a powerful force count, single camera power consumption is less than 1W.
R & D personnel in accordance with plans made for further optimization of this proposed algorithm, relying on self-developed AI chip in accordance with plans, the mere wearing, under the conditions of gait characteristics, has been able to do ReID accuracy of face recognition 2017-2018 . Such a high precision, not only accelerated the heavy pedestrian recognition of large-scale commercial floor, more to unlock a new application scenarios, to provide users with more comprehensive and ultimate experience.
2017, Apple FaceID face recognition as the representative of commercial applications began to spread across the globe. Today, brush face payments, car brush face has penetrated into our daily lives. Reasonable to expect a world-class ReID algorithm, coupled with self-developed AI chip in accordance with plans, the next field of computer vision industry looks forward to the "killer application" is coming.
