Abstract:

To address the challenges of significant geometric modeling errors, severe detail loss, and low training efficiency in neural radiance field(NeRF) reconstruction under defocused blurred input scenarios, this paper proposes two optimization strategies. One strategy is introducing Triplane features generated by the pre-trained LRM as prior knowledge, and combining a lightweight decoder and directional LoRA module to replace large MLP, thereby reducing parameters and shortening convergence time. The second strategy is integrating a differentiable blurring model into the volumetric rendering step. By jointly optimizing the radiation field and spatially variable blurring kernels, reconstruction accuracy under defocused blurred scenarios is enhanced. Simulation experiments show that the proposed model achieves significantly better reconstruction metrics than the original NeRF and comparative methods under defocused blurred scenarios, with shorter training time and higher-quality geometry and texture reconstruction. It can provide an effective solution for 3D reconstruction under blurred scenarios.

Key words: neural radiance field(NeRF), pre-trained Triplane, differentiable blur modeling, 3D reconstruction, defocused blur