Bridging 3D Editing and Geometry-Consistent Paired Dataset Creation for 2D Nighttime-to-Daytime Translation

Xiao Cao, Yuyang Zhao, Robby T. Tan, Zhiyong Huang

Keywords: Dataset creation, 3D editing, Generative mode, Nighttime-to-Daytime translation

Abstract:

Creating paired nighttime-to-daytime translation datasets remains a challenging and impractical task, as keeping every object static at different times is impossible. While 2D generative models can synthesize paired data for appearance and style translation, they often fail to maintain geometric consistency.
In this paper, we propose a novel paired synthetic dataset creation pipeline that leverages 3D editing techniques to convert daytime 3D datasets into nighttime degraded scenes, generating geometrically consistent high-quality image pairs.
Through this approach, we construct the first paired synthetic dataset for nighttime-to-daytime translation with geometric consistency.
The synthesized data pairs can effectively enhance nighttime-to-daytime editing performance of various 2D generative models both qualitatively and quantitatively, demonstrating the advantages of using 3D editing for paired synthetic visual dataset generation.
Code and Dataset are available at github.com/massyzs/3DEdting4Translation.git.

Submission Number: 1