Transfer Learning From Vision Transformers Or Convnets For 360-Degree Images Quality Assessment?

The watershed transform is a common step in different image processing tasks. With the fast development of general-purpose computing on GPUs, a number of parallel watershed algorithms have been introduced for that setup. We propose two novel parallel watershed algorithms: one intended for relatively larger images and another for smaller images. Our algorithms are based on the combination of the parallel path reduction technique, introduced in our previous paper on the topic, and the parallel version of the Union--Find algorithm. We show through multiple experiments, both in 2D and 3D, that our algorithms achieve unmatched execution times. On a typical gaming GPU, our algorithm processes an 800 megavoxel image in around 2.5 seconds.