Shulou(Shulou.com)06/02 Report--

Some time ago, there was a popular Mini Program "guess painting Song" on Wechat. You can draw simple strokes on your phone for machine recognition, and you can also compete with friends to see who is the soul painter.

In fact, guess the painting song is not the first release in China, as early as last November, Google launched a called "Quick,Draw!" The online game, that is, the original web page of "guess painting Little Song". In this game, players need to draw the required patterns within 20 seconds, such as simple strokes of objects such as basketball or TV.

Google "Quick,Draw!" The main difference between "guess painting Song" and "guess painting Song" lies in "Quick,Draw!" Is to use the mouse painting, and "guess painting song" can be directly on the screen with finger painting, if you are interested, you can click this link to enter the game: https://quickdraw.withgoogle.com/, feel the difficulty of drawing with the mouse, definitely much harder than finger painting.

Quick,Draw! Game home page

The game is so popular that it attracts many players to participate. Generous Google does not embezzle the wisdom of these netizens. Google has opened up the data set that players draw in the course of the game. This data set includes 345 categories, including TV, basketball, shorts and so on. More than 15 million players around the world have contributed nearly five million paintings.

Some categories contained in the quick draw dataset

Google makes this interesting data set available to the public, providing opportunities for all developers, researchers and artists to explore, learn and research.

After I got this interesting data set, I did some analysis and got some interesting results.

In the course of the game, players are required to finish painting in 20 seconds. In such a short time, what they draw is basically their own subconscious cognition of this object. I try to use the method of image stacking to explore the similarities and differences in describing different objects in different regions and countries.

Image stacking method

In the process of stacking, the number of pictures corresponding to different painting themes in different countries is different. I filtered less than 1000 pieces of data, and if there were more than 1000 pictures on a subject in a country, 1000 were randomly selected for stacking. The results are roughly as follows

France should be the country most familiar with the Eiffel Tower. If we compare the real Eiffel Tower with the pictures stacked above, we can see that the basic shape of the tower has been depicted, and there is a triangular gap at the bottom of the tower. It seems that in everyone's subconscious, the Alfie Tower is a triangle with a blank structure below.

The stacked images look interesting, with the feeling of peeling cocoons in the data. The effect of stacking is different for different themes, and the recognition degree of the stacked image is different according to the difference of topic complexity.

As the name implies, the category of five-pointed stars is the five horns, and people in all countries should have a unified understanding of the five-pointed stars, so the painted five-pointed stars are more or less the same.

Ice cream also exists widely all over the world. The classic ice cream structure should be egg peel below and ice cream pulled out by ice cream machine on top. The outline is very fixed and the structure is simple. There is basically no big difference between different people when they draw. But there are some categories of different people who draw will have big differences.

The structure of this category of tiger is very complex, and there are great differences in different people's understanding of the tiger. Some people may draw the tiger's head, and there may be a big "king" on it, while others draw the shape of the whole tiger. When stacked, the features are not very obvious.

In addition to the tiger, which is already a complex category, some categories are too simple to see their features when stacked. "River" is one of them.

Due to gravity, rock, altitude differences and other reasons, it is difficult to find a straight river in the world. In people's minds, the river is usually curved, so when different people paint, it is easy to describe the river in different places, and the curvature and width of the river are also quite different, which results in the low recognition of the stacked "river" pictures. Differences in stacked images

After the images were stacked, I was pleasantly surprised to find some interesting things. First of all, there is a certain preference in the image areas that people depict. Maybe Google takes the upper right as the origin when collecting graffiti trajectory data. Almost all the stacked images I observe are depicted more on the right and in the upper direction.

I have not found the answer to whether this phenomenon is caused by the way Google calculates the trajectory. But this phenomenon provides us with traversal of stacked pictures.

Graffiti towards

Although graffiti is a very simple line to describe the structure, but in the process of painting, different people will doodle in different directions, this phenomenon is like the three views in the university "engineering drawing".

Classic orthographic views

Chairs are common things in life, and graffiti has different structures from different angles. Here are four examples.

Graffiti on chairs in different countries

It can be seen that Australia and the United Arab Emirates are facing to the left, while Taiwan is facing to the right. The chairs painted in South Africa are regular, not biased on both sides, facing the middle!

The salient features of graffiti

In addition to the inconsistent orientation of objects such as chairs, I also found that people focus on the key features of objects in the process of graffiti. For example, when drawing crabs, the two large pliers are very obvious features, and almost all countries have thickened the crab pliers.

Crab this kind of big pincer feature should belong to the rough distinguishing feature, when drawing this kind of object, should focus on graffiti on the obvious feature, the system can quickly identify!

Local texture of graffiti

Texture information is very important for object recognition. The quick draw dataset contains many balls, such as basketball, football, baseball and so on. There are certain lines on the surface of these balls, so that people can tell whether it is basketball or baseball at a glance.

Baseball

These three kinds of balls are very popular sports all over the world, and their shapes are all spherical.

The texture of basketball is similar to that of cutting watermelons vertically, as is the case on almost all basketballs.

The texture of football is a typical combination of pentagons and hexagons. A football consists of 20 regular hexagons and 12 regular pentagons.

The structure of football is strikingly similar to that of C60. C60 is a carbon cluster, which consists of 60 carbon atoms to form a football-like 32-hedron, including 20 hexagons and 12 pentagons. When the 60 C atoms are arranged in space, they form a spatial arrangement position with the most stable chemical bond, which happens to be consistent with the arrangement of the football surface lattice.

Compared with the complex texture of basketball and football, the structure of baseball is much simpler.

There is usually a hard core inside the baseball, and the outer skin of the ball center is usually made of cowhide. Workers press the leather with an "8" mold with holes to get

A "leather coat", print the product number and factory date on the inside of the leather coat, and then apply glue on the inside of the leather coat. And then there's the last step in making baseball.

The worker glued and fastened two pieces of leather to the center of the ball and began to sew the ball! The hole left when pressing leather works at this time. Each baseball needs 108 stitches! Isn't that awesome!

After sewing the ball, just pass the ball through the press and print the logo, and a super handsome baseball can leave the factory.

When doodling these three kinds of balls with different surface texture complexity, the texture recognition is easy to be different. I found three countries "Philippines", "South Korea" and "United States" in the data set to observe the rules of graffiti on these three kinds of balls.

It can be seen that among the three kinds of balls, the order of recognition from high to low is "baseball" > "basketball" > "football". This result is basically consistent with the above analysis of the texture complexity of the three balls.

It seems that if you want to be identified as quickly as possible in quick draw games, texture features are also very critical!

There are a lot of interesting things about the quick draw dataset. I'll update this article later when I have time.

This article is published on Li Siyuan's blog "machines are learning".

Original link: http://www.siyuanblog.com/?p=1829

Welcome to scan the code and follow my official Wechat account: gather for the tower.

Series of blog posts are also posted in Huawei Cloud Community: https://bbs.huaweicloud.com/community/usersnew/id_1542351719401316

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.