Deep Digging into the Generalization of Self-Supervised Monocular Depth Estimation

Self-supervised monocular depth estimation has been widely studied recently. Most of the work focused on improving performance benchmark datasets, such as KITTI, but offered a few experiments generalization performance. In this paper, we investigate backbone networks (e.g., CNNs, Transformers, and CNN-Transformer hybrid models) toward estimation. We first evaluate state-of-the-art models diverse public which have never seen during network training. Next, effects texture-biased shape-biased representations using various texture-shifted datasets that generated. observe Transformers exhibit strong shape bias CNNs do texture-bias. also find show better for compared to models. Based these observations, newly design with multi-level adaptive feature fusion module, called MonoFormer. The intuition behind MonoFormer is increase by employing while compensating weak locality adaptively fusing representations. Extensive proposed method achieves datasets. Our shows best ability among competitive methods.