The front array brushed the screen of the NetEy cloud music brushing the song and sang the song, and everyone was turned.
"Upload a personal photo, AI will analyze your facial features and recommend you a song. Identify QR code is experienced"
Using Net Easy Ai's face key positioning technology, a scan of his own photos, while sweeping, launching the sugar shells:
"It's very cured, you will feel happy."
"Well, a pair of self-contracted romantic BGM eyes"
"This is a face that is a heart-moving face."
Hey, this emotional value of Max's small AI, the small sisters on the social media said it will not stop.
Although the love story is a little bit a little shoud, it shoulder "propagation positive energy, promoting the true sweet and beautiful" glorious mission, and there is still a scientific basis behind it.
01
Aesthetic basis:
"Beautiful ugly to look at the proportion, refuse to force handsome"
Although there is a difference in the aesthetic situation of people with social and popular cultures. For example, someone will create a new name "ugly" and "yin soft", this kind of "ugly" and "yin soft". Handsome "is created after the fans in the star personality buff, that is, the love of the lover, the eyes of Xi Shi. If you judge the beauty from the perspective of the art, there is not much change in the standard for thousands of years, that is, the ratio.
Everyone knows that Da Vinci is very respectful of gold segmentation (gold ratio), which is very good to practice this theory in his portrait. It is also the perfect face ratio of people, such as the aspect ratio of the face. The inner corner angle, the eye spacing is proportional, and the mouth is proportional to the dimension.
But this set of theory is too inconvenient, it is not conducive to spreading, the people can't have a big smell of people with a feet in order to judge the beauty. So a plastic surgeon combined with the proportion of more than ten years of insighting experience, and built a facial proportion of face that was said to be only the most beautiful talents - Marquet mask.
You don't think this is in this way, this mask has experienced a lot of verification. You can try to take the front of the star, and it will not deviate too much. For example, Zhang Ziyi, facial lines and facial features are seamlessly docked with this mask. In addition to international chapters, I also took other fronts of the stars in the eyes of ordinary people, and they were different.
"International Chapter and" Everyone Angel "Audrey's face, it is really a hundred people.
Image source Baidu picture
At this point, the proportion of Papath masks and Da Vinci's face constitutes a aesthetic basis for love.
But there is a bug here: people are moving. Why is almost all the shootings of all an anchor always placed on the side? Because this angle face looks the best, the face proportion is most matching. So this has another requirement for emotional AI: measurement of authenticity.
02
Data basis:
"Conscience AI, Tong Hao Neit"
In order to ensure the fair measurement of the whole people, this H5 calls NetEase Ai's face analysis interface, and is accurately identified by the positioning of the face-to-face key point, regardless of the photo being at any angle or an environment. Whether you are shooting
Still side shot,
Can accurately capture your facial data.
In addition, the H5 is easily reached in the release of the PV on the day of the release, which means that the love story is to identify a thousand-level face images within a day, and at least 10 million love will be said.
Of course, this is nothing, using distributed technology, NetEase Ai's daily advertising services can support 20 million days, now 10 million PV, can come back again.
03
The popularity and "bug" of a human face analysis
Many talents have realized that people's face analysis techniques actually have this effect, but in fact, the face analysis is not high-tech technology, iPhone X is unlocked, Alipay's face verification And each time Jacky Cheung concert police uncle can recognize fugitives through face-to-face, these are one of face analysis technology.
In the early years, Google uses this technology to make a brain loop clear application - see where your photos and museums are most like.
Identification effect can feel it. The upper left corner is matching, the lower right corner is the source of the painting. (Image source twitter)
First come to a set of low points.
A set of high scores.
I don't know if I don't know if I don't know, passing the gods.
Why do you have such a wonderful matching result?
To explain this problem, we must first know the two processes of face recognition: acquisition (feature extraction), match (model build + feature comparison).
Google's "People and Portraits" is to collect your photos first, match with the paintings of the museum; iPhone X will let you turn around and blink in the camera before the new machine is set, and the purpose is also to collect your face. Information, do matching when you unlock the screen.
Therefore, the accuracy of identification results is primarily subject to two factors.
First, collect accuracy,
Second, the size of the training man's facial library built for the model.
Explanation: You and Canon 5D beat it with the children's toy camera, the former shooting may not be recognized, this is "collecting accuracy"; but if you use the children's camera You finally formed a photo all day, your parents may have identified this is you from some "signing movements" (character). This is because "Training Man Face is very large".
So in order to make a good face recognition, the AI engineers need to expand the training man's faculty used for model builds while needing to collect people's face images with more advanced devices.
For example, iPhone X uses a TrueDepth camera system, a dual camera, one of which is an infrared camera, and also contains various sensors. It is the difference between the 2D face recognition is to use structural light, and the infrared spot of the shape information of the shape information is to the human face. The left camera collects the spot of the spot according to the variation and size, and each position is determined by one. Depth and direction information, finally get a point cloud generating 3D model. Such a result is to unlock your mobile phone with preventing your plane photos.
Image source stationmaster home
On the expansion of the trainer, for the protection of personal privacy information, the AI engineer usually uses the face picture data provided by volunteers (about 10,000 people), but it is impossible to know that this face information is too old. . However, don't worry, AI engineers will use data amplification technology to amplify these raw data, which can generally increase the data size and dozens of times even hundreds of times. Read the full article, the original title: the mysterious power behind the cloud music brush screen H5, it turned out to be a manual intelligence that would say "love"
Article Source: [Micro Signal: GH_70D0CCE81C74, WeChat public number: Netease artificial intelligence] Welcome to add attention! Please indicate the source of the article.
Our other product:
