Learning Continuous Exposure Value Representations for Single-Image HDR Reconstruction

Predicted exposure stack w/o corresponding GT Debevec's method Debevec's method (a) Continuous exposure value representation (CEVR) (b) Existing methods (c) Our continuous LDR stack benefits HDR reconstruction LDR stack w/ predefined EVs LDR stack w/ dense EVs Output HDR Output HDR Inverse CRF Inverse CRF Figure 1: Single-image HDR reconstruction from continuous LDR stack. (a) Continuous Exposure Value Representation (CEVR) generates LDR images with continuous exposure values (EV) without corresponding ground truth during training. (b) Existing methods build LDR stacks only with EVs covered by training data, which brings less visible details for Debevec's method [11] to estimate an accurate inverse camera response function (CRF), resulting in artifacts on HDR results. (c) Our CEVR model enriches the LDR stack by including additional LDR images with continuous and dense EVs (red frames), allowing Debevec's method to predict a more precise inverse CRF and reconstruct more visually pleasing HDR images.