May 2021
In May, we added the Dataset Editor to the portal, improved panorama accuracy, improved linking of products with price tags, and much more.
Dataset editor
A new type of dataset has appeared on the portal. It includes all photos, in which recognition errors were noted during the accuracy check.
The Editor will allow you to correct automatic recognition errors and train the error detector so that it does not repeat errors.
In the dataset itself, you can view each box, correct its size and location, change the label, or delete it altogether. When you select a box in the photo, the system will highlight it in the list.
The Editor can be found on the portal, in the Datasets section. Later, we plan to add a dataset for the shelves.
Panorama Accuracy
We wrote about the Panorama previously. This is an algorithm that creates one large panoramic shot from several photos of a scene. This month we improved its accuracy.
The way it works
There are 2 photographs in a scene that are not related to each other in any way.
As a result, the system glued them into one frame. This was because the ceiling and floor look the same in the photo.
In order for the panorama to be glued without such errors, we need to find key points in two photos, and algorithmically correctly match them .
For this, a special map is used that displays matches. From the set of lines, the system looks for those that will satisfy the conditions. The most relevant points are taken from the points on the lines and a transformation matrix is built on them. This is necessary to correctly adjust one photo to match another.
The more accurately such points are found, the better is the picture quality on the gluing seam.
In addition, there were situations when the top of the photo was glued correctly, but the goods themselves on the shelf were glued incorrectly.
Now points are taken into account only inside the annotations. Now the system does not look at the top part with the banner, but only at what is inside the annotation.
Linking products with price tags
Vision algorithms see boxes of goods and price tags. But in order to tell how much which product costs, the system needs to make an assumption of which price tag belongs to which product.
We have completely rewritten the old algorithm. Now the linking of goods with price tags is determined more correctly and accurately.
Viewing the algorithm
You can see how the algorithm works in a photo scene. To do this, press on a photo and select Explain price.
