A lot of modern phones today have background blur features. Whether they have a depth camera or not, and they estimate everything from just a 2D image, very often we are faced with various artifacts, the most common problem being hair. Company Apple he wants to combat this with a new artificial intelligence system Depth Pro. She recently published a scientific article on this topic, where she described her new system, its training and its testing and results.
These are really very good. Although he did not win in all categories in estimating the distance, he was the best overall by a margin. In estimating the edges of objects and the focal length of the lens, it won in all parameters and tests examined. In the test images, it could be seen that it lacks, for example, typical artifacts at the edges of objects. It also works very well with hair, or for test images with fur.
It’s also up to two orders of magnitude faster than competing methods, but that doesn’t mean it can get by on weak hardware. It took pictures with a resolution of 2.25 MPx (1536×1536 pixels) in 0.3 seconds, while it was tested on an older Nvidia V100 computing card (it was introduced in 2017). It has 5120 CUDA cores and its theoretical performance in FP32 is roughly at the level of the GeForce RTX 4060, but the performance in FP16, texture rate (GTX/s) and pixel rate (GPx/s) is more at the performance level of the GeForce RTX 4070 Ti.
Source: www.svethardware.cz
