Lu, Xiaohu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin
In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 859-867. 2020.
We present a novel method for generating panoramic street-view images which are geometrically consistent with a given satellite image. Different from existing approaches that completely rely on a deep learning architecture to generalize cross-view image distributions, our approach explicitly loops in the geometric configuration of the ground objects based on the satellite views, such that the produced ground view synthesis preserves the geometric shape and the semantics of the scene. In particular, we propose a neural network with a geo-transformation layer that turns predicted ground-height values from the satellite view to a ground view while retaining the physical satellite-to-ground relation. Our results show that the synthesized image retains well-articulated and authentic geometric shapes, as well as texture richness of the street-view in various scenarios. Both qualitative and quantitative results demonstrate that our method compares favorably to other state-of-theart approaches that lack geometric consistency.
We presented a novel approach for satellite-to-ground cross-view synthesis. In particular, we proposed an endto-end trainable pipeline that takes a single satellite image and generates a geometrically consistent panoramic RGB image. We thus proposed a neural network with a differentiable geo-transformation layer that links a semantically labeled satellite depth image with a corresponding semantic panoramic street-view depth image which is finally used for photo-realistic street-view images generation. The geometric consistency across images significantly improves the accuracy of the skyline in the panoramic ground view which is especially important for urban areas. Our experiments demonstrate that our method outperforms existing approaches and is able to synthesize more realistic street-view panoramic images and in larger variability
If you use the code, please cite:
@inproceedings{lu2020geometry,
title={Geometry-aware satellite-to-ground image synthesis for urban areas},
author={Lu, Xiaohu and Li, Zuoyue and Cui, Zhaopeng and Oswald, Martin R and Pollefeys, Marc and Qin, Rongjun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={859--867},
year={2020}
}