Abstract: Some implementations involve a process that identifies a subset of points in an image and creates descriptors for these points. The detection and descriptor process may use one or more neural networks. In some implementations, the process includes a neural network that uses one or more fixed (e.g., weight independent) neural network layers to perform certain functions that can be performed more accurately and/or efficiently than via non-fixed (e.g., weight-based) layers. In some implementations, for example, a neural network includes a layer that determines orientation formulaically within the neural network. Such orientations may be determined convolutionally (e.g., using sliding patches) but are not determined based on internal node weights within the layer that were determined during the training of the neural network.

Abstract: Training a descriptor network includes obtaining a first image of a scene from an image capture device, applying a sensor sensitivity model to the first image to obtain shadow-invariant image data for the first image, and selecting a first patch from the image. Training a descriptor network also includes obtaining a subset of the shadow-invariant image data corresponding to the first patch, and training the descriptor network to provide localization data based on the first patch and the subset of the shadow-invariant image data.

Abstract: Providing localization data includes obtaining a first image of a scene associated with a first condition, determining one or more target conditions, and applying an appearance transfer network to the first image to obtain one or more synthesized images comprising the scene, wherein the scene is associated with the one or more target conditions in the synthesized image. A first patch is selected from the first image, wherein the first patch comprises a keypoint, and an image location is determined for the first patch. Then one or more additional patches can be obtained using the synthesized images and the image location. A descriptor network may be trained to provide localization data based on the first patch and the one or more additional patches.