What is the next outlet after face recognition?
On this issue, the industry seems to have a consensus. Since the face recognition capability of AI has surpassed that of human beings, the attention of academia and industry has gradually turned to another subject of more scientific research significance and application value-Pedestrian Re-identification (ReID).
Recently, Yitu Technology has made a new breakthrough in the field of ReID, refreshing the current best results (SOTA) of the three authoritative data sets in the industry, and the algorithm performance has reached the industry's highest standard so far, greatly expanding the boundaries of algorithms and applications.
Note: The YITU algorithm does not use spatio-temporal information, does not perform re-ranking and re-ranking, and obtains
under such restrictions. Remember that YITU entered intelligent voice at the end of 2018 , And immediately set a new record of recognition accuracy in the field of Chinese speech recognition. The world's first cloud vision AI chip was launched in May 2019, and it is "commercially released". It seems that no matter which technology field it enters, Yitu can quickly push the overall industry level to a new high and accelerate the industrialization of technology.
What is the key behind this?
Pedestrian re-identification (ReID), the "killer application" after face recognition
In actual scenes such as transportation, industrial manufacturing and urban planning, 99% of the images are not The human face or the part of the human face is extremely blurred, only a few pixels in size, and the role of face recognition at this time is relatively limited.
Pedestrian re-identification (ReID, also known as "pedestrian re-identification"), refers to the retrieval of pedestrians under a network of multiple camera equipment, using more comprehensive information such as gait movements and physical characteristics to identify people, regardless of Use alone or in combination with face recognition can exert greater application value.
In addition to the frequently mentioned application scenarios such as smart retail, smart transportation, and smart cities, the application of ReID technology will also make daily life more convenient: amusement parks are easier to find lost children, pets/home robots can accurately identify their owners from their backs Or customers and provide corresponding services.
However, because ReID needs to find the same person in images or videos shot by different cameras, and the areas covered by these cameras do not overlap with each other, resulting in a lack of coherent information, and the postures of the characters in different pictures, The behavior and even the appearance (such as being upright, sideways, and back) will change greatly. The lighting, background and occlusions of the scene will be different at different times (there are often other people with similar body shapes and clothes in the background). The resolution is also high and low, and the position of the characters in the picture is far and advanced. These all pose a great challenge to the ReID technology.
Deeply optimize the ReID algorithm framework, AutoML replaces manual algorithm tuning
With its own engineering and R&D capabilities, Yitu Technology has deeply optimized the ReID algorithm framework and significantly improved the algorithm efficiency. By combining AutoML Other cutting-edge technologies have further innovatively realized the automatic search and iteration of model parameters, breaking through the traditional algorithm development process that relies on manual design and tuning by algorithm researchers, reducing labor costs and making the algorithm more generalized.
This time Yitu’s self-developed algorithm will be used in the industry’s three most influential ReID data sets Market1501, DukeMTMC-ReID, CUHK03, and will measure the performance of the algorithm, the two key indicators "Rank- 1 Accuracy) and "Mean Average Precision" (mAP) 6 items have all been improved, which fully demonstrates Yitu's technical strength and further stabilizes the leading position of the Chinese technical team in this task.
It needs to be pointed out that the high hit rate in the first place only means that the algorithm can accurately find the easiest to identify or match among many images, and it does not reflect the true ability of the model, especially the performance of dealing with complex scenes.
Therefore, the mAP value needs to be combined when evaluating the performance of the ReID algorithm.It reflects the comprehensive retrieval performance of the system. The higher the mAP value, the better the practicability of the system, which can be checked both fully and accurately, and can better cope with multiple occlusions, low light, and blurred pictures.
Algorithm + computing power to accelerate the commercialization of ReID
Facing another industry record, the Yitu team was very calm. According to Yitu’s R&D staff, this review is just an attempt. The scale and complexity of the ReID project that Yitu has implemented in the industry has far exceeded the three major data sets. It can be said that the existing ReID benchmarks in academia have Cannot reflect the highest level of industry algorithms.
For example, Market-1501 is collected in Tsinghua University. Pedestrians (ID) are basically Asians wearing short sleeves, shorts and skirts. DukeMTMC-reID is collected in Duke University. ID is mainly body For Europeans and Americans wearing winter clothes, these data collected in specific scenes and specific time periods are often inconsistent with the distribution of images in the real world. In real scenarios, the ReID algorithm needs to be able to perform high-precision and rapid recognition across time periods, across scenes, and across image acquisition devices with different imaging qualities. The data distribution is far more complicated than the existing academic data set.
These realistic factors have caused the existing ReID academic data set to be unable to effectively simulate or restore the actual situation. Therefore, the benchmark based on the existing ReID data set has great limitations. According to Yitu researchers, the industry needs a better ReID data set and a more comprehensive algorithm to measure the data set, at least for commercialized algorithms.
The ReID task in the actual combat scenario not only puts forward higher requirements on the algorithm, but also requires a more efficient chip to provide powerful computing power support. The lack of either of the two will affect the actual application value of ReID. At present, Yitu is a company with both algorithm and computing capabilities. Yitu invested in the research and development of the cloud AI chip QuestCore (Quest) in 2017, and "released for commercial use" in May 2019. QuestCore is the world's first cloud vision AI chip, providing powerful computing power, and the power consumption of a single camera is less than 1W.
In the actual application of ReID, Yitu's R&D personnel have further optimized the algorithm proposed this time. Relying on Yitu's self-developed AI chip, it has been able to achieve ReID under the condition of only wearing and physical characteristics. The accuracy of face recognition from 2017 to 2018. This not only accelerates the large-scale commercialization of ReID, but also unlocks new application scenarios.
In 2017, the commercial application of face recognition represented by Apple's FaceID began to spread globally. Nowadays, face-swiping payment and face-swiping rides have penetrated into our daily lives. There is reason to believe that the world-class ReID algorithm, coupled with self-developed AI chips, the next "killer application" in the field of computer vision that the industry is looking forward to has arrived.
雷锋网雷锋网
