BP-NeRF: End-to-End Neural Radiance Fields for Sparse Images Without Camera Pose in Complex Scenes

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Novel view synthesis from sparse image sequences is challenging, particularly without known camera poses.
• Accurate depth priors and camera motion constraints are crucial for improving synthesis quality.

Purpose of the Study:

• To propose an end-to-end network, BP-NeRF, for synthesizing high-quality novel views from sparse image sequences, even without camera poses.
• To enhance accuracy by integrating reliable depth information and robust camera motion estimation.

Main Methods:

• Developed the RDP-Net module to generate accurate depth maps and assess their quality for sparse image sequences, providing a depth prior.
• Constructed a novel loss function based on 2D-3D feature consistency to improve camera pose estimation accuracy and robustness.

Main Results:

• BP-NeRF effectively estimates camera motion trajectories and generates novel view images from sparse indoor and outdoor complex scenes.
• Experimental results on LLFF and Tanks datasets demonstrate superior performance compared to existing methods in novel view synthesis without camera poses.

Conclusions:

• BP-NeRF offers a robust solution for novel view synthesis from sparse image data, overcoming limitations of insufficient features and unknown camera poses.
• The proposed method achieves state-of-the-art results, highlighting the importance of accurate depth priors and geometric consistency in camera motion estimation.