Realistic Pedestrian Shadow Generation by 2D-to-3D Object-Lifting

Yu-Sheng Su, Jen-Jee Chen, Yu-Chee Tseng

Abstract: This paper studies the shadow generation problem in an outdoor street scene. Previous methods only rely on GAN-based model with the sun position to reconstruct the street view. Lacking in related dataset, we propose an evaluation method to make sure of the effectiveness. Our model applies a 2D-to-3D lifting method, casts the 3D object with sun estimation, and finally merges the shadow with a predicted sun brightness. Limited with lifting objects, our model casts more precise shadows than GAN-based model. Tuning with the brightness, the cast shadow will be in consistence with the whole street view. Then in our evaluation, we demonstrate better performance over the state-of-the-art models by 0.009 in LPIPS. Therefore, casting with a 3D human object is a feasible solution for shadow generation in the future.