FlexHDR: Learning Adaptive Exposure Control for Split-Pixel HDR Imaging

Split-pixel sensors offer a promising single-shot solution for High Dynamic Range (HDR) imaging, especially in safety-critical applications like autonomous driving. However, conventional auto-exposure methods typically use shared exposure settings across subframes, overlooking their distinct sensitivities and noise characteristics and thereby limiting dynamic range utilization and downstream perception performance. In this work, we propose FlexHDR, a task-driven neural exposure control framework that incorporates sensor priors to predict exposure parameters for individual subframes. Our method combines a differentiable image formation model for closed-loop exposure optimization during training with a cross-attention module that fuses subframes in feature space for downstream perception. To support reproducible evaluation, we develop a calibrated split-pixel camera simulator and a synthetic HDR scene radiance dataset covering diverse automotive scenes. Experiments demonstrate consistent improvements in object detection and instance segmentation over existing auto-exposure and tone-mapping methods under challenging automotive HDR conditions. These results highlight the benefits of independent per-subframe exposure control and end-to-end training for robust HDR perception.