Unifying Guided and Unguided Outdoor Image Synthesis

Muhammad Usman Rafique, Yu Zhang, Benjamin Brodie, Nathan Jacobs; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 776-785


Given a source image, our goal is to synthesize novel images of the same scene under different conditions, which could include changes in the time of day, season, or weather conditions. We consider two variants, unguided and guided synthesis, both of which require a way to generate diverse output images that cover the range of possible conditions. For the former task, the layout of the output image should match the source image and the conditions should appear realistic. For the latter task, the conditions should match those of a provided auxiliary guidance image. We address both tasks simultaneously using a probabilistic formulation, with separate distributions for each task, and use an end-to-end training method. We draw samples from these distributions to synthesize plausible images of the source scene. We prepare a new large-scale dataset and propose three benchmark tasks. The dataset, benchmarks, and evaluation code are available at https://mvrl.github.io/un_guided.

Related Material

[pdf] [supp]
@InProceedings{Rafique_2021_CVPR, author = {Rafique, Muhammad Usman and Zhang, Yu and Brodie, Benjamin and Jacobs, Nathan}, title = {Unifying Guided and Unguided Outdoor Image Synthesis}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2021}, pages = {776-785} }