Selective detection of visual cues for automated assistants

Abstract

Techniques are described herein for reducing false positives in vision sensor-equipped assistant devices. In various implementations, initial image frame(s) may be obtained from vision sensor(s) of an assistant device and analyzed to classify a particular region of the initial image frames as being likely to contain visual noise. Subsequent image frame(s) obtained from the vision sensor(s) may then be analyzed to detect actionable user-provided visual cue(s), in a manner that reduces or eliminates false positives. In some implementations, no analysis may be performed on the particular region of the subsequent image frame(s). Additionally or alternatively, in some implementations, a first candidate visual cue detected within the particular region may be weighted less heavily than a second candidate visual cue detected elsewhere in the one or more subsequent image frames. An automated assistant may then take responsive action based on the detected actionable visual cue(s).