Today we’re talking about another winner of our CVPR competition: the Herbarium 2021-Half-Earth Challenge, a story that ended up being so accurate that people suspected we’d hired a botanist to compete.
Previous review: If you have seen the story of the last champion Shao Lin, “After only 9 days to win the CVPR competition, the family dog changed his schedule”, today’s first year champion has two main differences: first, he began to prepare 3 months in advance; Second, he took part in three challenges at once.
The unusual path of violent youth
Leo’s first year, in the eyes of his teammates is a “violent youth.”
Walk on the road if there are stones in the way, others choose to detour, he chose to push; Play table tennis together, no matter what kind of tricky ball the other side, a violent smash to eat all over the world; As part of the team outing, we were walking on a steady mountain road, but he insisted on running to the wild hillside to open his own way, and never looked back without seeing the cliff…
“It’s intense, it’s radical, but it’s also very reassuring.”
The young Leo, along with his virgo, Aquarius and Capricorn colleagues, won three prizes in the CVPR 2021 competition, taking home the first, second and third place prizes. (The algorithm team that believes in astrology)
A group of straight men forced “grass” : almost green
Yan Baoming (flower name: 1st year), who studied microelectronics as an undergraduate and was recommended to Peking University to study for a PhD in physics, was originally a layman in computer science.
“I didn’t want to weld circuits, and I thought basic physics was too difficult to see a breakthrough in a short time, so I taught myself the algorithm.”
Even if it is so terrible learning ability, and so fierce speech, the first year also never thought that the competition itself to engage in is – botany.
Herbarium 2021-Half-Earth Challenge is a technical field of “Large-scale instance level object recognition”. The data set of the Herbarium 2021-Half-Earth Challenge is 2.5m plant sample images of the 6.5w class collected from several large botanical gardens in The Americas, Oceania and other Half of the world, and is used to train plant recognition algorithms. Assist botanists in plant identification, discovery and protection of new species.
Different from the identification and judgment of Taobao commodities in the first year of the work, the classification and category of plants are quite detailed. It requires professional knowledge to distinguish different plants, and it is very difficult to distinguish visually. In addition, the sample of the data set is extremely unbalanced, with only three samples for the smallest category. How to improve the accuracy of the long tail category is the core problem of this competition.
“When we visualized some of the results, people were like, ‘What’s the difference? !” (For those algorithmic straight guys who can’t even read the color of their lipstick)
The competition team led in the first year participated in 3 different CVPR competitions this time. When his teammate collapses under a pile of green pictures, he advises him that we should give it up and put more energy into other races. In the first year, “Violent youth” resolutely refused.
In his mind, the algorithm itself is to quickly debug the problem so that you can find the next one. They were on their way to submitting 38 results.
The geek spirit of the techie: Play three at once without fear
Originally for the “instance level object recognition” this field, the first year has its own pride and confidence.
During the business of taobao live broadcasting hotspot, the requirement is to correlate the goods with the broadcast room, and the user clicks on the goods to jump to the broadcast room that is being explained. Tao department multimedia algorithm team in the field of retrieving commodities has a very good plan, before the relevant field competition also has a relatively high data indicators. When they ran this race with a similar scheme, they were less than 70 per cent accurate.
“Classification is not the same as retrieval,” he said. “What we found is that you have to decouple the learning of features from the classification. It’s this idea that gives us the dark technology of this competition.”
This is where the first year of “Lion” brings the greatest encouragement to friends. No matter how desperate a situation may seem, he is always coming up with new ideas and strategies, and more on the spot.
“If you want to eat early, don’t let the first year talk at the meeting.” Teammate LAN æž» can’t laugh or cry.
Director Jiyu was also deeply touched by the endless energy and breakthrough spirit in the first year of 2018. When he was recruited in the first year of 2018, Jiyu directly threw away an important business demand and told him to go online in 2 months. In the first year, he completed the engineering problems and algorithm deployment in one and a half months without any work experience. At the beginning of 2019, he asked to show some technical depth, and in the first year, he published two papers as a writer. Now, when he thinks, why not play a game? So in the first year, he led his team to win the championship, second place and third place in three CVPR courses.
The other two competitions of the first year group are the Hotel-ID 2021-Hotel Recognition Challenge and the Differentiated Expressions from Videos (EEV) Challenge. All of us have just graduated a few years ago. We have to ensure that we do not delay the business at hand, but also promote the three competitions at the same time, which poses a greater challenge for us to manage our time and energy during the competition.
If the machine could talk, it would probably feel it the most and want to jump up and hit someone.
“Because of the limited resources, our principle is that we must not let the machine idle, run after one to the next, when we run on a new task, when we sleep. Weekends are no exception.”
In the first year, the team divided the priorities of the three competitions. Each partner had his own key entries and independent follow-up, and maintained the cooperative operation mode of “scheme attempt – problem summary – communication improvement”. In the first year, we kept the research on the progress of each competition and the next step.
“My buddies are really good, like Len æž», our training logs are all on OSS, and he’s crazy enough to send all his logs to his phone so he can scan the results in real time. Even if you see something different in the data at 2 or 3 in the morning, you pop up and grab the computer. That’s the geek spirit of us techies.”
The winner of the Half-Earth Challenge was 2.2% ahead of the runner-up, and the other two were second and third, respectively. This is the first time that the team attempted instance-level retrieval in the visual image dimension in the first year. In the future, he hopes to expand to the larger category of content retrieval. On the one hand, he can better understand users’ retrieval intention through multi-modal information, and on the other hand, he can achieve better multi-dimensional representation of video content by combining multi-modal information. In the process of exploring and solving many basic problems such as cross-modal learning and causal knowledge decoupling, we will bring better content consumption experience to Taobao users with technological breakthroughs.
How to learn “superpowers” across fields?
Four years ago, he secretly frozen fruits with 200 degrees below zero liquid nitrogen next to the instruments of the cryoquantum laboratory of Peking University. He never thought that he would make computer vision algorithms in the future.
Physics is a relatively “boring” and great subject, requiring long hours of experimentation and refinement. In the first year, I watched him at the beginning of his admission. From the time he started to build instruments, until he graduated from the doctoral program, his teachers and younger sisters were still working on this problem… He began to think about how he could find a way to realize his life value with faster feedback.
“My idol is Iron Man Musk, who has made disruptive achievements in space exploration, automobiles, energy, high-speed rail, Internet, artificial intelligence and other fields.”
Like his idol, The key to the first year, who primed himself on “physics in computers, physics in computers”, was the ability to “transfer learning”.
Musk’s answer, published in a Reddit AMA, previously described a two-step process for achieving this superhuman domain connectivity. First, learn to deconstruct knowledge into a number of basic principles, just like a large tree to ensure understanding of the trunk; Second, refactor these fundamentals into new areas, constantly asking yourself two questions: “What does it make me think of” and “why does it make me think of it”.
When we learn to think about the common points at a more abstract level, we can better understand that genius is not innate, but more scientifically correct learning skills.
Of course, the first year like to go hiking and mountaineering. The same is physics department of hardcore wife, also can accompany him to drive Tibet, in the vast world between mountains and rivers, the thought is broken or recombination, it is not a matter.
It’s okay. “High Achievers” is also the biggest loser at Texas Hold ’em
Surprisingly, compared to his Virgo, Aquarius and Capricorn teammates, the lion is more of a money giver.
“He’s just, if he starts to see that he has a good hand, he suddenly raises a lot of chips. They all contribute to us in the end.” Good gay friend LAN æž» as a conservative, can not help but ridicule.
To this, the first year solemnly said: to play easy to go to the game, must not think so much, go to the right. (Ok, we bought it.)
Different from the impression that students with good grades “high scores and high energy are good at what they learn and play”, or “high scores and low energy are nothing but good grades”, the first year belongs to the third kind of permeable: it is good to have fun, and it is good to play. He is not devoted to thinking and skills to “play” to the extreme level, nor is he clumsy and inflexible, but also bears social mental pressure.
Play is the essence, the talent, the joy and the foundation of being human. It is also the cradle of passion, creativity and imagination. Whether serious or casual, as long as maintaining a relaxed and happy state of mind, get the spirit of relaxation and pleasure, to be worthy of their own life.
Life is evergreen, fun is Paramount, work and rest.