Reconstructing 3D scenes from multi-view images has been a longstanding challenge in computer vision, where neural radiance fields (NeRFs) have shown significant promise in generating realistic renderings of novel viewpoints. However, most existing NeRF approaches rely on either accurate camera pose information, a large number of input images, or both. Reconstructing NeRFs from a limited number of views without known camera poses is particularly challenging and poses an ill-posed problem.

To solve the problem, a research team led by Xin WEN published their new research on 15 October 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.


The team proposed CAD-NeRF, a method capable of reconstructing 3D scenes from fewer than 10 images without any known camera poses. Specifically, they constructed a mini library of CAD models from ShapeNet, rendering these models from various random viewpoints. For sparse-view input images, the method retrieves a CAD model with a similar shape from the library, which is then used for density supervision and pose initialization. To address the new challenge of avoiding pose conflicts among views in uncalibrated NeRF methods, they introduced a novel multi-view pose retrieval technique. The object’s geometry is then trained with guidance from the CAD model, while the deformation of the density field and the camera poses are optimized jointly. Texture and density fields are further refined through fine-tuning. All stages of training are conducted in a self-supervised manner. Comprehensive evaluations on both synthetic and real-world datasets demonstrate that CAD-NeRF effectively learns accurate density fields with significant deformations from the retrieved CAD models, showcasing strong generalization capabilities.

Future work could focus on overcoming the limitation of requiring image order information, expanding the method’s applicability to a wider range of scenarios, and developing more efficient techniques for accurate pose retrieval.
DOI:10.1007/s11704-024-40417-7