Continuous training of an object detection and classification model for varying environmental conditions

Abstract

Apparatuses, systems, and techniques for continuous training of an object detection/classification model are provided. A first image depicting an environment is identified. Data for an object detected in the first image is obtained based on outputs of a machine learning (ML) model. A determination is made of whether a level of confidence that the object corresponds to an object class satisfies a confidence criterion. If so, first conditions corresponding to the environment depicted in the first image are identified. Noise characteristics associated with a second image depicting the environment are determined based on a difference between the first conditions and second conditions of the second image. The first image is augmented based on the noise characteristics to generate a third image depicting the environment based on the second conditions and the detected object. Training data associated with the third image is provided to train the ML model.