The document presents a novel multispectral transfer network (MTN) for unsupervised depth estimation that operates effectively under varying lighting conditions throughout the day and night. It discusses the architecture of MTN, which utilizes multi-task learning to predict depth, surface normals, and semantic labels from thermal and RGB inputs, along with specific components like interleaver modules and adaptive scaled sigmoid functions for enhanced training stability. Experimental results demonstrate MTN's competitive performance in both day and night scenarios compared to traditional depth estimation methods.