Multi-Scale Implicit Surface Reconstruction for Outdoor Scenes

The images used to reconstruct 3D models in outdoor scenes are generally captured at different scales. Accurately reconstructing geometry from multi-scale images has not been extensively addressed. This work proposes a new volume rendering method that combines cone sampling and implicit surface representation to better model geometries from multi-scale images. Besides, to address the problem of unsmooth gradient of different frequency bands in general position encoding, we propose a dynamic position encoding strategy. Meanwhile, we propose an adaptive sampling strategy in image space to focus on the important regions near the reconstructed surfaces and the difficult regions where rendering error exists. Experiments are conducted on the Tanks and Temples dataset as well as the aerial photography dataset. The results show that, compared to the state-of-the-art methods, our work can produce competitive or better high-quality surface reconstruction, especially for scenes with multi-scale images and complex geometric structure.