Image Reconstruction from Local Descriptors Using Conditional Adversarial Networks

Many applications rely on the local descriptors extracted around a collection of interest points. Recently, the security of local descriptors has been attracting increasing attention. In this paper, we study the possibility of image reconstruction from these descriptors, and propose a coarse-to-fine framework for the image reconstruction. By resorting to our gradually reconstructing network architecture, the novel multiscale feature map generation algorithm, and the strategically designed loss functions, our proposed algorithm can recover the images with very high perceptual quality, even partial descriptors are provided only. Extensive experimental results are reported to show its superiority over the existing algorithms. Our study implies that the local descriptors contain surprisingly rich information of the original image. Users should pay more attention to sensitive information leakage when using local descriptors.