Dynamic image region selection for visual inference

摘要

A region (e.g., a portion) of a full-size image is selected for analysis by a vision inference model. The region of the image may be dynamically selected for the image based on environmental characteristics of the image. These environmental characteristics may include semantic information about the environment that may be determined from a previous image of the environment, road structure information, control information, and motion information. These environmental characteristics may thus describe expected areas of interest in the image. The selected region of the full-size image is resized to a size compatible with a vision inference model which identifies semantic information about the current image. This semantic information may then be used to select a region of a subsequent image.