"???" Meng Fanqi walked and put the contract documents into his bag. As soon as he opened the door, he heard this sentence.
He froze on the spot and took a closer look at the person in front of him. He was about half a head taller than himself, and he was very tall.
He was wearing a loose shirt and sweatshirt, with obvious signs of training. At this moment, he was staring at himself with a smile, and his body movements seemed a little awkward.
"Are you...?" Meng Fanqi raised his left eyebrow and asked tentatively.
"Wang Kai, brother." Wang Kai chuckled and asked Meng Fanqi to sit down. "Please sit down, please sit down. I am specifically responsible for handing over the code."
Wang Kai said, pulling his laptop between the two of them, "The internal account and password have been created for you. You need to upload the entire project to the designated location, and then have some necessary code comments, introduction and running methods. process.
I have to make sure that others can understand the code in each part, so the last task today is for me to understand the entire algorithm. After going through the code and technology, the handover is complete. "
"No problem, I understand this." Meng Fanqi was very familiar with code handover, version control with internal tools, what specifications the code has, where comments must be written, and what the rules and restrictions are on the naming of variables and functions.
Not to mention, he also worked at Shang Hai's Baidu, and his code habits will definitely not make mistakes that do not meet the requirements.
He spent more than a week re-optimizing the code. The reason why it took so long was largely because the handover required a lot of necessary code comments and document writing.
So that others can understand quickly.
"Wang Kaige, right? I know a lot about these things. I've written all the notes and documents." Meng Fanqi said, doing some operations on Wang Kaige's computer, and pulled the latest version of the code from his github private repository. to local.
"I'll simply go through the entire process with you. I definitely won't be able to figure out all the details today. Let's leave a contact information. We will definitely need to communicate many times in the future."
"It's Wang Kai, brother." Wang Kai has a habit of speaking. When he sees powerful people, he can't help but call him brother. In line with his burly figure, there is some strange contrast. "Mr. Yu and I have the same name, and we both have the same name as Kai."
"Oh oh oh." Meng Fanqi suddenly realized, "Aren't you all master's degree holders? You must be at least 25 or 26, right? What do you call me brother?"
"This year is twenty-four, brother." Wang Kai replied, "There are specialties in the arts, and the code is the same. The one who masters it is the brother. In our group, I am the younger brother among the younger brothers. Looking around, everyone is my brother. .”
"Does it mean you are the best?" When the three senior executives were there just now, Meng Fanqi always felt like a leader was there, which he was not used to. Now that I am joking with ordinary employees like Wang Kai, it feels more like working.
"Then why were you sent here?"
"It must have been a stupid bird that flew first." Wang Kai was not ashamed of his talent, but rather proud of it. "If I can understand the handover, it means there will be no problem at all. Everyone in the entire institute will definitely be able to understand it." Easy to take."
"Still planning, still planning, come and listen to the explanation from elder brother." Seeing that Meng Fanqi had already pulled down the project and opened the documents and code, Wang Kai quickly stood up and greeted him.
Amid the energetic greetings, more than a dozen research members came and directly surrounded the two of them.
"Sister-in-law, sit here!" Seeing Han Ci standing and watching, Wang Kai suddenly felt that it was time for him to show off, and quickly stood up and gave up his seat.
"Don't talk nonsense. How could I have such a good life?" Meng Fanqi thought this boy was quite interesting. "Let's get down to business. I have given you the documents and the draft of the paper. Please read this carefully. I will not directly said."
"Today I will mainly talk about the main context and logic. The YOLO detection algorithm means you only look once. It means that the image is scanned only once, and its main speed improvement comes from this.
The process of a classic detection method is as follows: first analyze the picture through computer graphics (or deep learning) method, find several areas where objects may exist, crop these areas, and put them into an image classification in the machine, classified by the classifier.
The stupidest thing to do is to repeatedly scan the image with multiple size marquees, and then use some methods to filter out the most likely areas. These methods are two-step. Look at the picture once to select the area of interest, and look at the picture again to classify this area.
My approach is to directly divide the entire picture into multiple uniform grids, each of which can output the category of the object and the coordinates of the detection frame. Which area the object falls in is responsible for predicting the object.
My network output is one-dimensional. The first N numbers represent the probabilities of N categories, the following X numbers represent the probabilities of X detection frames, and the last 4X numbers are used to describe the position of the detection frame on the image. "
"How can it be one or two hundred times faster?" A quick-thinking Baidu researcher asked aloud, "The traditional method that pays more attention to speed will reduce the number of proposed regions, usually no more than a thousand at most. According to your statement , you need at least dozens of grids, and each grid requires several detection frames, so the difference doesn’t feel very big.”
"If we use a 7x7 grid + two detection frames, the total number will be only 98 detection frames. In addition, there are also great adjustments in the design of the backbone network and the specific operation sequence of non-maximum suppression. space."
"The performance should be greatly affected, right?"
“How should we achieve multiple types of goals?”
"In this case, the detection of small targets should be a problem, right?"
As Meng Fanqi explained, the researchers present had a deeper and deeper understanding and had more and more questions.
Meng Fanqi has to say that Robin Li's institute has indeed gathered a group of outstanding young programmers with jumping and creative thinking.
After only listening for half an hour, I was already able to ask quite in-depth and meaningful questions.
Among them are the sources of speed improvement, parameter settings and meanings, which Meng Fanqi explained very well.
There are also concerns about small target detection, which is indeed a flaw in the YOLO algorithm.
When discussing technical content, time always flies by very quickly. Unknowingly, two or three hours have passed, and Meng Fanqi has finally completed all the tasks for this trip.
"Let's go. I signed a big deal today. It's going to take a lot longer than I thought." Such high-intensity negotiations and technical discussions were indeed a bit too much for him mentally. "What do you mean, can I treat you to a meal?"
"Go to the city." Han Ci suggested with a smile.
In fact, for most girls, it is difficult to have any good feelings in their hearts after this trip. It lasted for several hours, and it would be so boring to come here like this.
However, for Han Ci, this kind of talent and dedication in the professional field is a bonus.
