Palasts: Behind the scenes. Computer Vision

Have you ever wondered how Google knows which plant is on the photo you’ve just made? How does Apple differentiate you from your friend on all the selfies and can create a film with all the food photos of last half a year? How does Palasts’ algorithm find similar products and knows what is gonna be a good fit to your interior? In this blog post we want to guide you behind the scenes of how Computer Vision algorithms “see” images and make sense of it.

Unlike humans, computers process all the information in the numeric form. For example, images are seen by a computer as a matrix of numbers. Each of those numbers encode the color of the corresponding pixel.

Processing these numbers, computer vision algorithms can compare different images, detect various furniture objects and even make interior recommendations.

But how exactly can a computer make sense of what is in the image?

Here is where Machine Learning comes into play. Instead of comparing two dimensional matrices of numbers directly, the goal is to teach an algorithm how to transforms every image into a vector, so that each number reflects some important characteristic of an image.

In order to teach an algorithm to make sense of images we have to show it thousands of examples of different objects with different characteristics. Like teaching a small baby the difference between red and green pencil, we have to show the algorithm examples of furniture, ask to classify it, give feedback and repeat until it learned to give mostly right answers. Trained that way, the algorithm can learn, for example, how to compare images, identify similar ones, or classify images into different groups.

Let’s have a closer look into classification. 

In this example, the first number aims to answer the question if there is a table on that image. Learning from many thousands of examples, Palasts’ algorithm knows that tables look differently in 98% cases and says that this image could contain a table with 2% probability.  It also suggest the image to be a sofa with the highest probability. 

This model and example is enhanced to recognize different objects, colours, styles, shapes, contrasts etc. Hereby we can identify the clients interior design taste in a very implicit way and help the client find what he/she really wants and needs without high effort or cost.

 

Artificial intelligenceInterior designMachine learning

Kommentar hinterlassen

Alle Kommentare werden von einem Moderator vor der Veröffentlichung überprüft