Noise augmentation
AssigneeNVIDIA Corporation
InventorWangren Xu
Life of this patent
Estimating the term…
Abstract
Disclosed are apparatuses, systems, and techniques for testing and training hardware devices and computer vision models using images augmented with noise. In one embodiment, the techniques include augmenting a plurality of images with noise to generate a plurality of sets of noise-augmented images (NAIs). Individual sets of NAIs include noise of a respective noise level, applying one or more sets of NAIs of the plurality of sets of NAIs to determine one or more accuracy metrics characterizing performance of a computer vision (CV) model for one or more noise levels of respective one or more sets of NAIs, and publishing the CV model in association with the one or more accuracy metrics.
Claims19
A method comprising:
- augmenting a plurality of images with noise to generate a plurality of sets of one or more noise-augmented images (NAIs), wherein an individual set of one or more NAIs comprises noise of a respective noise level of a plurality of noise levels;
- applying the plurality of sets of one or more NAIs to determine a mapping of the plurality of noise levels to a plurality of sets of one or more accuracy metrics, each set of the one or more accuracy metrics characterizing performance of a computer vision (CV) model for a respective set of the plurality of sets of one or more NAIs;
- creating a CV model bundle comprising the CV model and the determined mapping;
- causing the CV model bundle to be provided to one or more users of the CV model;
- receiving a request for a noise-tolerant CV model, the request comprising a target noise tolerance level;
- and
- retraining, responsive to the received request, the CV model using additional one or more sets of one or more NAIs to obtain the noise-tolerant CV model, wherein at least one set of one or more NAIs of the additional one or more sets of one or more NAIs has the noise level at or above the target noise tolerance level.
A system comprising:
- a memory device to store a plurality of images;
- and
- a processing device communicatively coupled to the memory device, the processing device to:
- augment the plurality of images with noise to generate a plurality of sets of noise-augmented images (NAIs), wherein an individual set of NAIs comprises noise of a respective noise level of a plurality of noise levels;
- apply the plurality of sets of NAIs to determine a mapping of the plurality of noise levels to a plurality of sets of one or more accuracy metrics, each set of the one or more accuracy metrics characterizing performance of a computer vision (CV) model for a respective set of the plurality of sets of NAIs;
- create a CV model bundle comprising the CV model and the determined mapping;
- and
- cause the CV model bundle to be provided to one or more users of the CV model.
A system comprising:
- one or more sensors to:
- obtain a plurality of images;
- and
- one or more processing units to:
- augment the plurality of images with noise to generate a plurality of sets of noise-augmented images (NAIs), wherein an individual set of NAIs comprises noise of a respective noise level of a plurality of noise levels;
- apply the plurality of sets of NAIs to determine one or more accuracy metrics characterizing performance of a computer vision (CV) model for one or more noise levels of respective one or more sets of NAIs;
- publish the CV model in association with the one or more accuracy metrics;
- receive a request for a noise-tolerant CV model, the request comprising a target noise tolerance level;
- and
- retrain, responsive to the received request, the CV model using additional one or more sets of one or more NAIs to obtain the noise-tolerant CV model.
Dependencies are inferred from the claim text.