A Novel Nighttime Close-up and Far-away Reflection Dataset with Absorption Effects

U.-In Chan, Tsung-Jung Liu, Kuan-Hsien Liu, Chun-An Lin

Abstract: Existing methods for synthesizing reflection removal datasets meticulously account for various factors, including diverse glass properties (e.g., thickness, color, direction), the processing of reflection images (e.g., Gaussian noise, smoothness, intensity), and the weighting between transmission (T) and reflection (R) images. However, these approaches overlook the distinction between day and night reflections and the significant influence of the distance between the photographer and the glass on the reflection image. Therefore, we propose to incorporate considerations of distance, lighting conditions (day and night), and combined absorption effects into the synthetic dataset generation process to produce a reflection dataset that closely resembles real-world scenarios. Moreover, creating a nighttime reflection dataset poses greater challenges than its daytime counterpart. We aim to demonstrate the applicability and address the challenges of our proposed synthetic dataset through an approach that integrates absorption effects.