Prior is Power: High Dynamic Radiance Field with Depth Optimization

Current 3D reconstruction-based HDR image synthesis methods often suffer from depth reconstruction errors due to limited scene views and overfitting to training data. These errors manifest as incorrect depth perception, discontinuous surfaces, and overfitted patterns. To address these limitations, we propose a novel depth-optimized 3D reconstruction algorithm for HDR imaging. Our approach leverages a high-dynamic imaging algorithm based on dynamic scene 3D reconstruction, incorporating monocular depth priors to guide depth optimization. A monocular depth estimation model is employed to provide additional depth information during training. Depth ranking supervision is further introduced to refine the model’s depth optimization process. A dual-path ray sampling pipeline, combining random ray sampling and pair-wise ray sampling, enables simultaneous depth ranking and color supervision. Experimental results demonstrate that our proposed algorithm significantly improves scene depth reconstruction without compromising image quality.