Image Reconstruction based on Local feature Descriptors

In this project, we show that an image can be reconstructed using local descriptors, with or without complete geometrical metadata. We use greedy algorithms to progressively learn the missing information before reconstruction and colorization is performed. Our experiments show that most of the vital information about a query image can be recovered even if scale metadata is missing. Compared to images reconstructed with scale information, we find that there is no significant decline in image quality, and a close resemblance of the original image can be obtained via colorization as a post-processing step.