Technology: Autograph
Today, more than 230,000 photos have appeared on Yandex.Maps for approximately 130,000 homes - more than the number of all buildings in Moscow, which is multi-million. Of course, so many houses are beyond the power of photographers. At least it would be a titanic work. Photos are taken with the help of a specially developed technology based on Panorama -Avtoprofotoma.
Street panoramas consist of a multitude of photographs depicting thousands of buildings. But viewing all the panoramas and “manually” cutting pictures of houses from them is difficult and time consuming. Panoramas are taken every 20-50 meters, and only 60 thousand panoramas were made for Petersburg alone. Therefore, it was necessary to develop a mechanism that would automatically find in the panoramas of the building, cut out and select the best shots. This mechanism is called Autophotograph.
The autophotograph finds in the panoramas of the house, comparing the panoramic image and the corresponding fragment of the usual map. Knowing the coordinates of the house on the map and its outlines, Avtoprofoto determines how the house is located relative to the point of shooting. To do this, it calculates the azimuth for the extreme points of the building (the angle between the direction to the north and the object). And then, using these azimuths, he finds a house in a panoramic picture.
')
First, an Avtoprofoto takes pictures of all the houses in a row, regardless of whether they are clearly visible. It may be that the house in the picture closes the foliage or unsuccessfully parked truck. Such photos, of course, will not help users to view the building. Therefore, the next auto photograph assesses the quality of images. For this, a special classifier was created, which analyzes the photo and puts its mark - from one to three.
When evaluating pictures, the classifier takes into account a number of factors: the linear dimensions of the building, the shooting angle, the distance from the survey point to the house. These parameters help to determine in what perspective and how large the building is shot. In addition, the image itself is analyzed: color saturation, color and geometric components. So, if the picture contains many multidirectional lines, most likely, the house is poorly distinguishable. And if the photo has a lot of green or yellow-orange color, the building may be covered with foliage.
The qualifier-evaluator was configured using machine learning . To teach the car to determine the quality of images, experts manually selected several thousand good photos. After analyzing the images, the classifier identified the most important factors for determining the quality and calculated their significance.
Typically, the building falls on several panoramas at once, so the Avtoprofotogras get 15-20 shots of one house. Among them, you need to choose some good photos taken from different points. It is convenient for users when they can see the building from different sides. To do this, Avtofotograf divides all the pictures into groups (by sector) - depending on what point and from what distance they were taken. And then selects some of the best photos from different sectors.
The selected images are uploaded to the Yandex.Foto service, and from there they go to Yandex.Maps. These photos and see users on the building card.
Street panoramas consist of a multitude of photographs depicting thousands of buildings. But viewing all the panoramas and “manually” cutting pictures of houses from them is difficult and time consuming. Panoramas are taken every 20-50 meters, and only 60 thousand panoramas were made for Petersburg alone. Therefore, it was necessary to develop a mechanism that would automatically find in the panoramas of the building, cut out and select the best shots. This mechanism is called Autophotograph.
The autophotograph finds in the panoramas of the house, comparing the panoramic image and the corresponding fragment of the usual map. Knowing the coordinates of the house on the map and its outlines, Avtoprofoto determines how the house is located relative to the point of shooting. To do this, it calculates the azimuth for the extreme points of the building (the angle between the direction to the north and the object). And then, using these azimuths, he finds a house in a panoramic picture.
')
First, an Avtoprofoto takes pictures of all the houses in a row, regardless of whether they are clearly visible. It may be that the house in the picture closes the foliage or unsuccessfully parked truck. Such photos, of course, will not help users to view the building. Therefore, the next auto photograph assesses the quality of images. For this, a special classifier was created, which analyzes the photo and puts its mark - from one to three.
When evaluating pictures, the classifier takes into account a number of factors: the linear dimensions of the building, the shooting angle, the distance from the survey point to the house. These parameters help to determine in what perspective and how large the building is shot. In addition, the image itself is analyzed: color saturation, color and geometric components. So, if the picture contains many multidirectional lines, most likely, the house is poorly distinguishable. And if the photo has a lot of green or yellow-orange color, the building may be covered with foliage.
The qualifier-evaluator was configured using machine learning . To teach the car to determine the quality of images, experts manually selected several thousand good photos. After analyzing the images, the classifier identified the most important factors for determining the quality and calculated their significance.
Typically, the building falls on several panoramas at once, so the Avtoprofotogras get 15-20 shots of one house. Among them, you need to choose some good photos taken from different points. It is convenient for users when they can see the building from different sides. To do this, Avtofotograf divides all the pictures into groups (by sector) - depending on what point and from what distance they were taken. And then selects some of the best photos from different sectors.
The selected images are uploaded to the Yandex.Foto service, and from there they go to Yandex.Maps. These photos and see users on the building card.
Source: https://habr.com/ru/post/141063/
All Articles