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: In one implementation, a method includes: obtaining an input image, wherein the input image is captured by an image sensor having a rotational orientation with respect to a direction of gravity; obtaining a gravity direction estimation associated with the rotational orientation of the sensor; generating, from the input image, a rotationally preprocessed input image by applying one or more transformations to the input image based on the gravity direction estimation; providing the rotationally preprocessed input image to the machine learning sub-system; and identifying, using the machine learning sub-system, a visual feature within the rotationally preprocessed input image.