High Dynamic Range Imaging from RAW Domain with a New Benchmark Dataset and a Multi-scale Network

High dynamic range (HDR) imaging aims to generate realistic HDR images from multiple low dynamic range (LDR) images with different exposures. Conventional HDR reconstruction pipeline often contains two steps, which first converts the camera-recorded RAW data to SRGB data using the image signal processor (ISP) operations and then fuses the multi-exposure images to produce the HDR image. However, the computation cost imposed by ISP operations becomes larger as the number of input images increases. In addition, the converted sRGB images are typically degraded by lossy ISP operations, which increases the difficulty of HDR reconstruction. In this work, towards high-quality HDR image reconstruction, we propose a pipeline to perform HDR reconstruction directly from RAW domain, which takes a set of multi-exposure RAW LDR images as input and outputs a high-quality RAW HDR image. Considering the lack of RAW HDR reconstruction datasets, we first construct a benchmark dataset, which enables training and evaluating HDR reconstruction methods. Then, we propose an end-to-end multiscale network for HDR reconstruction, which performs alignment and fusion in a coarse-to-fine manner. Specifically, we devise a multi-scale deformable alignment module (MS-DAM) for aligning the RAW LDR inputs and a progressive fusion module (PFM) for progressively fusing the aligned inputs. With our proposed MSDAM and PFM, the useful information in input LDR images can be properly utilized to help reconstruct the final HDR image. Extensive experiments demonstrate the effectiveness and superiority of our method and dataset. Our codes and dataset will be available at https://github.com/syujung/RAW-HDR.