Abstract: A deep learning network can perform three-dimensional (3D) image reconstruction of a scene from multi-view calibrated two-dimensional (2D) images. The network can include a convolutional neural network that performs feature extraction to generate feature pyramids corresponding to features at different levels of resolution. The feature pyramids can be used to compute a respective cost volume for each feature pyramid at each level of resolution, with the cost volume incorporating a learnable parameter that corresponds to a weight allocated to the “reference” feature pyramid relative to other feature pyramids. A depth map for each input image can be generated based at least in part on the cost volume.

Abstract: A deep learning network can perform three-dimensional (3D) image reconstruction of a scene from multi-view calibrated two-dimensional (2D) images. The network can include a convolutional neural network that performs feature extraction to generate feature pyramids corresponding to features at different levels of resolution. The feature pyramids can be used to compute a respective cost volume for each feature pyramid at each level of resolution, with the cost volume incorporating a learnable parameter that corresponds to a weight allocated to the “reference” feature pyramid relative to other feature pyramids. A depth map for each input image can be generated based at least in part on the cost volume.