Enhancing Self-Supervised Monocular Depth Estimation with Self-Distillation

Monocular depth estimation aims to recover the depth information of each pixel in an image using a sequence of monocular images. Compared to supervised learning methods that rely on large amounts of labeled data, monocular depth estimation has gained significant attention due to its low cost and greater flexibility. In recent years, with the rise of unsupervised learning methods, significant progress has been made in monocular depth estimation. However, since monocular images lack ground truth depth labels and rely on geometric relationships between adjacent frames to infer depth, the model struggles with insufficient supervisory signals, making accurate depth estimation challenging, especially in complex scenes. To address this lack of supervisory signals, this paper proposes a self-distillation method based on data augmentation, aimed at providing additional supervision to improve depth prediction accuracy. During training, data augmentation is applied to the input images, generating corresponding depth maps as pseudo-labels. The pseudo-labels are then used to supplement the supervision through self-distillation loss. Unlike traditional student-teacher models, which typically rely on multi-stage knowledge transfer through complex architectures for self-distillation, our method maintains a simplified network structure. Additionally, the use of data augmentation enriches the training data, allowing the model to learn more diverse features. Experimental results on the KITTI dataset demonstrate that the proposed method achieves competitive performance, validating the effectiveness of using self-distillation to supplement supervisory signals.