In recent years, do all kinds of kissing scenes in TV dramas make people blush and heartbeat after watching them? But what you don't know is that a new profession, kissing drama appraiser, is derived from it. Do you know what is the main job of a kissing judge? Let's get to know.
'at the beginning, my girlfriend and I said that my job is to train the computer to recognize the kissing scenes. She thought it was incredible after listening. Who can give you money to do this?' Ji Hengshan graduated from Beijing University of Aeronautics and Astronautics in 2016. His master's degree is computer vision. Generally speaking, the purpose of this major is to 'teach the computer to understand the world'. After graduation, he first worked as a programmer in an Internet company. Last September, he joined the Alibaba technology team and became a senior algorithm engineer of Youku. His main job was to teach computers to recognize kissing scenes, also known as "kissing drama appraiser".
'this post is just in line with my major. It can convey happiness and happiness to you. I can watch online dramas in the name of work. I like it very much. 'talking about her work, Ji Hengshan smiles. He usually likes watching online dramas, most like the military division alliance, which is Wu Xiubo's iron powder.
In the past decade, China's online video content consumers have reached nearly 600 million. "More and more people are using their mobile phones and computers to watch online dramas, movies and sports competitions. Every time they go to work and take the subway, they can see someone with their mobile phones and earplugs watching videos. 'Ji Hengshan found that everyone's need is no longer simply to watch the drama, but also to watch more exciting and interesting clips and collections, and more actively find the stars or scenes they want to watch.
To achieve search and intelligent editing, we need to train the machine to find out the corresponding scenes from the massive materials and label them, just like picking out crabs and turtles from the vast ocean. The "kissing play" of the male and female protagonists conveys beauty and happiness, which is generally the climax of the plot development, so it becomes the most common label in the drama. How to teach the machine to recognize kissing has become the primary problem for Ji Hengshan and his colleagues.
This seemingly cold AI black technology really needs a lot of 'artificial' in the initial stage. Jihengshan cut out thousands of pictures from the picture library and video, and marked "kissing" pictures one by one. This process is called 'data tagging'.
The 'kissing appraisers' in the team printed out pictures of' kissing 'from different angles and pasted them on the wall. A group of people were discussing how to calculate' kissing '. Supply map
'hundreds of kissing images should be marked in a day. At most, thousands of kissing images should be marked. Staring at the computer screen for a day, I feel a little numb at last. 'generally, the more accurate and quantitative the data is, the better the effect of the model will be,' said Ji. Naturally, the effect of machine recognition will be better. For example, in order to make a beautiful picture, the gesture of "kissing" is also very different. Some of them look at each other affectionately, while some of them are far away. The "kissing drama appraisers" in the team have to print out the pictures of "kissing" with different angles and stick them on the wall. A group of people have a heated discussion about how to calculate "kissing".
"Kissing opera appraiser" is a technical job, which requires tedious later debugging according to the response of the machine. The machine learns the features of thousands of pictures. Then it can recognize any picture of kissing. So far, Ji and his colleagues have labeled more than 20000 "kissing scenes".
Jihengshan's team currently has 8 people. With the rapid expansion of business, the "kiss drama appraisers" also began to contact and study other scenes, such as crying drama and playing drama. Next, they also need to identify shooting, throwing, shooting, celebration and other pictures in sports games, so that the website can edit and produce a wonderful collection of games. They not only need to see pictures, but also often 'visit' to watch and experience the small differences of each scene and action in the competition or program.
At present, 5-minute video full link processing can be completed in 10 seconds. For example, the 50 minute video of the play "three lives, three lives and ten li peach blossom" is cut into 3000 pictures. The machine can pick out Yang Mi and Zhao Yating's kissing drama in five seconds.
In recent years, there have been some new phenomena in the entertainment industry, such as the stars are more and more alike, how to train machines to recognize gulinaza and dirieba, how to distinguish the "three little ones" of tfboys, and how to make trial and error and carry out model training repeatedly. Distinguishing "star face" from "red face" twins "has become a new challenge for Ji and her colleagues.