Abstract: Disclosed implementations include a depth generation method using a novel teacher-student GAN architecture (TS-GAN) to generate depth images for 2-D images, such as RGB images, where no corresponding depth information is available. An example model consists of two components, a teacher and a student. The teacher consists of a fully convolutional encoder-decoder network as a generator along with a fully convolution classification network as the discriminator. The generator takes 2-D images as inputs and aims to output the corresponding depth images. The teacher learns an initial latent mapping between 2-dimensional and co-registered depth images and the student applies the latent mapping to provide feedback to the classification network for refinement.

Abstract: Disclosed implementations include a method, apparatus and computer media for learning an optimal graph in the form of a tree topology defining a sequence that can be used by a learning network for image recognition. Image data representing the image of an object is received and N landmarks are detected on the image using a deep regression algorithm, wherein N is an integer. A weighted, fully connected, graph is constructed from the landmarks by assigning initial weights for the landmarks randomly. An optimized tree structure is determined based on the initial weights. A sequence is generated by traversing nodes of the tree structure and a series of embeddings representing the object image are generated based on the sequence. The embeddings can be processed by a neural network to generate an image recognition signal based on the embeddings.