Speaker
Description
We present a class-agnostic approach to robotic gaze control using Fast Normalized Cut and DINOv3 visual features. The method partitions feature maps without explicitly constructing the pairwise similarity matrix defined by feature dot products, achieving linear scaling with the number of image patches per iteration. Recursive bipartitioning with a feature-similarity stopping criterion produces object masks without predefined object categories. These masks guide the robot’s head movements to keep an object centered in its field of view. The system runs in real time on a gaming notebook with a GPU, demonstrating its applicability to interactive robotic perception. It also provides object descriptors that support distinguishing between observed objects. This way, the robot can see objects of arbitrary classes.