Abstract: Embodiments of the present invention provide systems, methods, and computer storage media directed to optimizing the accuracy of local feature detection in a variety of physical environments. Homographic adaptation for facilitating personalization of local feature models to specific target environments is formulated in a bilevel optimization framework instead of relying on conventional randomization techniques. Models for extraction of local image features can be adapted according to homography transformations that are determined to be most relevant or optimal for a user's target environment.

Abstract: A method of training a 3D structure detector to detect 3D structure in 3D structure representation, the method comprising the following steps: receiving, at a trainable 3D structure detector, a set of training inputs, each training input comprising at least one 3D structure representation; the 3D structure detector determining, for each training input, a set of predicted 3D objects for the at least one 3D structure representation of that training input; and training the 3D structure detector to optimize a cost function, wherein the cost function penalizes deviation from an expected geometric relationship between the set of predicted 3D objects determined for each training in put.

Abstract: The present disclosure pertains generally to image feature extraction. Both transfer-learning and multi-task training approaches are considered. In one example, a machine learning model is trained to perform a geographic classification task of distinguishing between images captured in different geographic regions based on their visual content. In another example, a machine learning model is trained to perform an order recognition task of determining information about the order of an image sequence based on its visual content, where the order of the image sequence may be different than the order in which its constituent images were captured. A further example combines the two approaches. The knowledge gained by the ML model in learning one or more such tasks can be applied to a desired image recognition task, such as image segmentation, structure detection or image classification, e.g. with a pre-training/fine-tuning framework or a multi-task learning framework.

Abstract: Embodiments of the present invention provide systems, methods, and computer storage media directed to optimizing the accuracy of local feature detection in a variety of physical environments. Homographic adaptation for facilitating personalization of local feature models to specific target environments is formulated in a bilevel optimization framework instead of relying on conventional randomization techniques. Models for extraction of local image features can be adapted according to homography transformations that are determined to be most relevant or optimal for a user's target environment.

Abstract: The present disclosure pertains generally to image feature extraction. Both transfer-learning and multi-task training approaches are considered. In one example, a machine learning model is trained to perform a geographic classification task of distinguishing between images captured in different geographic regions based on their visual content. In another example, a machine learning model is trained to perform an order recognition task of determining information about the order of an image sequence based on its visual content, where the order of the image sequence may be different than the order in which its constituent images were captured. A further example combines the two approaches. The knowledge gained by the ML model in learning one or more such tasks can be applied to a desired image recognition task, such as image segmentation, structure detection or image classification, e.g. with a pre-training/fine-tuning framework or a multi-task learning framework.