SYSTEM AND METHOD USING PYRAMIDAL AND UNIQUENESS MATCHING PRIORS FOR IDENTIFYING CORRESPONDENCES BETWEEN IMAGES

A method of unsupervised neural network training for learning of local image descriptors is provided. A pair of images depicting a same scene is obtained. The pair of images includes a first image with a first pixel grid and a second image with a second pixel grid, wherein the first pixel grid differs from the second pixel grid. A neural network with an initial set of parameters is applied to the first image and the second image to generate feature maps F1 and F2. Each feature map comprises a grid of local image descriptors. An initial correlation volume C0 is determined based on F1 and F2. Based on C0, a high-level correlation volume CL is determined by iterative pyramid construction. CL comprises aggregated high-level correlations between iteratively constructed high-level patches of the first and second pixel grids. A uniqueness matching loss for F1 and F2 is determined based on CL. The neural network is trained by minimizing a loss function based on the uniqueness matching loss to generate an optimized set of parameters, thereby generating a trained neural network for determining optimal local image descriptors.