End-to-End Deep HDR Imaging with Large Foreground Motions


In state-of-the-art deep HDR imaging such as Kalantari’s, the problem is formulated as an image composition problem, by first aligning input images using optical flows which are still error-prone due to occlusion and large motions. In our end-to-end approach, HDR imaging is formulated as an image translation problem and no optical flows are used. Moreover, our simple translation network can automatically hallucinate plausible HDR details in the presence of total occlusion, saturation and under-exposure, which are otherwise almost impossible to recover by conventional optimization approaches. We perform extensive ualitative and quantitative comparisons to show that our end-to-end HDR approach produces excellent results where color artifacts and geometry distortion are significantly reduced compared with existing state-of-the-art methods.

European Conference on Computer Vision (ECCV), 2018