Combine Neural Radiance Fields in Different Reference Coordinates

Neural Radiance Fields (NeRF) leverage deep learning networks to restore high-quality implicit representations of 3D scenes from 2D images, surpassing traditional texture mapping in rendering quality. Typically, reconstructed images by NeRF exist in a reference coordinate system, necessitating consideration of each surface's information in a single shot. Overlapping parts in adjacent images are addressed by obtaining the camera pose through sparse COLMAP reconstruction. However, occluded regions, caused by contact with support surfaces, pose challenges for complete reconstruction. To address this, we propose a scene fusion algorithm based on NeRF, facilitating the amalgamation of neural radiance fields under different reference coordinate systems. This reduces data requirements and shooting conditions, enhancing the implicit 3D reconstruction of real objects. By estimating the translation and rotation relationship between image capture angles and the NeRF model, we fuse neural radiance fields in distinct reference coordinate systems, completing object surface information. Evaluation is conducted on both NeRF synthetic dataset and real objects.