LEARNING TO RECONSTRUCT 3D SHAPES BY RENDERING MANY 3D VIEWS

Methods, systems, and apparatus for obtaining first image features derived from an image of an object, providing the first image features to a three-dimensional estimator neural network, and obtaining, from the three-dimensional estimator neural network, data specifying an estimated three-dimensional shape and texture based on the first image features. The estimated three-dimensional shape and texture are provided to a three-dimensional rendering engine, and a plurality of three-dimensional views of the object are generated by the three-dimensional rendering engine based on the estimated three-dimensional shape and texture. The plurality of three-dimensional views are provided to the object recognition engine, and second image features derived from the plurality of three-dimensional views are obtained from the object recognition engine. A loss is computed based at least on the first and second image features, and the three-dimensional estimator neural network is trained based at least on the computed loss.