[11:08 Sun,7.August 2022 by Thomas Richter] |
A team of researchers from ETH Zurich and TU Munich, among others, has published a new deep-learning algorithm that extracts extremely precise objects from photos, i.e. can automatically separate (aka crop) an image object from the background. The neural network for dichotomous image segmentation was trained using 5,470 high-resolution (e.g. 2K, 4K or larger) images of various objects in front of different backgrounds, which were more or less distinct from the background. ;) Peacock cut out
The nice thing about this algorithm is that you can already try it yourself online with your own images using an unofficial demo of Dichotomous Image Segmentation huggingface.co/spaces/ECCV2022/dis-background-removal. However, the researchers point out that the currently released DIS Deep Model (DIS V1.0) was trained using a dataset with only a few images of animals, people, cars, etc., and therefore may not work as well with such objects. However, the team plans to release another version (DIS V2.0) soon for general use and testing, which will then cover a much wider range of object categories. ;) Neural network architecture The practicality of such automatic extraction of objects from photos for various tasks in image processing, 3D modeling, animation, or AR/VR is obvious: it saves work that would otherwise have to be done by hand, because the new method works more accurately (even for complex objects) than other current algorithms that attempt similar things. In the  paper you can read how exactly the new method works and also implement it yourself using the published program code.  more infos at bei github.comdeutsche Version dieser Seite: Neue KI stellt Objekte aus Photos automatisch perfekt frei |
 |
|
|