Self-Calibrated Neural Implicit 3D Reconstruction

Neural implicit representations have become the new trend in 3D reconstruction from multi-view images. Prior arts unify the implicit surface models and radiance fields, which enables high quality 3D reconstruction and eliminates the need for input masks. Unfortunately, such methods require both intrinsics and extrinsics for every input images to convert pixel coordinate points into spatial coordinate points. To this end, we propose a novel self-calibrated framework to directly reconstruct 3d structures from multi-view input images only. In industrial settings, our approach removes the necessity for real-time measurement and storage of camera external parameters and the need for computationally intensive tools like COLMAP for camera calibration. Instead of using additional network structures to predict the poses corresponding to the images, we set them as learnable network parameters and add global constraints together with local constraints as regularizers. The global regularizer explicitly constrains the relative poses between different views so that the learned poses conform to the global ge-ometric structure. While the local regularizer optimizes the local geometry of the surface in adjacent views and further improves the reconstruction quality. Experimental results demonstrate the superiority of our method when the need for calibrated camera parameters is removed.