Smart, Sparse Contours to Represent and Edit Images
share
We study the problem of reconstructing an image from information stored at sparse contour locations. Existing contour-based image reconstruction methods struggle to balance contour sparsity and reconstruction fidelity. Therefore, denser contours are needed to capture subtle texture information even though contours were not meant for textures. We propose a novel image representation where image content is characterized by contours with gradient information via an encoder-decoder network, while image details are modeled by a conditional generative adversarial network. We show that high-quality reconstructions with high fidelity to the source image can be obtained from extremely sparse input, e.g., comprising less than 6 details and fine structures in regions where no input information is provided. The semantic knowledge encoded into our model and the sparsity of the input allows using contours as an intuitive interface for semantically-aware image manipulation: local edits in contour domain such as scaling, translation and erasing, translate to long-range and coherent changes in the pixel space. Experiments on a variety of datasets verify the versatility and convenience afforded by our models.
READ FULL TEXT