Object labeling in images using dense depth maps

Abstract

There is provided a method for annotating digital images for training a machine learning model, comprising: generating, from digital images and a plurality of dense depth maps, each associated with one of the digital images, an aligned three-dimensional stacked scene representation of a scene, where the digital images are captured by sensor(s) at the scene, and where each point in the three-dimensional stacked scene is associated with a stability score indicative of a likelihood the point is associated with a static object of the scene, removing from the three-dimensional stacked scene unstable points to produce a static three-dimensional stacked scene, detecting in at least one of the digital images static object(s) according to the static three-dimensional stacked scene, and classifying and annotating the static object(s). The machine learning model may be trained on the images annotated with a ground truth of the static object(s).