Neodev

Take a look at the image next to this text. Do you see a bird? Do you see a bird standing on a wall? Do you see a bird standing on a wall and looking out over a harbor?

Object detection

Even if one option is claimed, the other observations are also correct. Many would probably first say that it is a bird (or perhaps even a seagull?) in the image. These people can likely easily draw a line around the object to indicate which pixels constitute the bird and which are something else.

It may seem obvious, but when letting a computer figure out which part of the image is of interest, it becomes a much more difficult problem. Shouldn’t the boats in the background be equally prominent objects as the bird?

Boni

Boni receives a large number of images showing products with various backgrounds. As the number of images grew, the need arose to automatically remove the background. Boni offers a service where users can select images of furniture and other interior details to easily visualize how a room could be furnished. Each part can be easily swapped out until the optimal combination is found and then ordered through them. In their case, it is crucial for each product image to be isolated for the visualization to be realistic.

Object detection algorithms have improved significantly in recent years, but they were not sufficient in this case. Often, parts of the product were cut off or parts of the background remained. A pre-trained model developed for “prominent object detection” was used to create an improved model using examples of Boni’s product images. The enhanced model is now used to isolate the suppliers’ images, and customers can even upload their own products to be isolated on the website.