Recursive Multi-Exposure Alignment with Spatiotemporal Decoupling for Efficient Burst HDR and Restoration

Tianheng Qiu, Qi Wu, Yuchun Dong, Shenglin Ding, Xuan Huang, Hu Wei, Guanghua Pan; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 1038-1047

Computing high dynamic range (HDR) RGB output from multi-frame low dynamic range (LDR) RAW input is a challenging task because it requires solving multiple subtasks including multi-frame fusion of different exposures, image restoration including denoising, deblurring, HDR imaging, and modeling RAW to RGB mapping. Solving the problem using a unified model is more difficult as these tasks need to be considered simultaneously.In this paper, in order to construct a generalized efficient Burst HDR and Restoration method, we propose the Recursive Multi-Exposure Alignment with Spatiotemporal Decoupling (RASD) algorithm. Specifically, in order to address the information discrepancy between multi-exposure data, we propose a recursive flow-guided alignment module based on multi-exposure alignment, which is used to provide more accurate multi-frame alignment. In addition, we introduce a spatiotemporal decoupling strategy to train the alignment and restoration tasks in stages to prevent possible optimization conflicts introduced between multiple tasks. Extensive experiments show that our proposed method obtains state-of-the-art performance, and we are the winner in the NTIRE 2025 Efficient Burst HDR and Restoration Challenge.