Exploiting the Potential of Self-Supervised Monocular Depth Estimation via Patch-Based Self-Distillation

Abstract

Perceiving scene depth and 3-D structure is one of the key tasks for Internet of Video Things (IoVT) devices to understand and interact with the environment. Self-supervised monocular depth estimation has demonstrated significant potential in leveraging large-scale unlabeled datasets to achieve competitive performance, thereby playing an increasingly important role in depth estimation. Despite recent methods providing additional supervisory signals through self-distillation strategies to improve depth estimation, an effective method for generating pseudo-depth labels suitable for addressing occlusion issues among elements far from the camera remains unexplored. To address this limitation, we propose a patch-based self-distillation learning framework to exploit the potential of self-supervised monocular depth estimation in recovering fine-grained scene depth. In the proposed framework, elements far from the camera within the input image are enlarged by enlarging and cropping operations in the patch-based self-distillation branch. Guided by photometric consistency, the model learns the detailed occlusion relationships among elements from the enlarged patches, producing patch depth maps with fine structures. In the main branch, which takes full-scale images as input, patch depth maps serve as pseudo-depth labels through self-distillation loss to provide additional supervisory signals for regions where photometric consistency fails to offer effective supervision. This forces the depth estimation network to recover fine structures of elements far from the camera in full-scale input images. Regarding the architecture of the depth estimation network, we introduce a bin-center prediction. In this prediction, a global aggregator based on self-attention provides additional scene structure queries for adaptive scene depth discretization. Finally, to encourage the model to explore more general cues for depth inference beyond road plane cues, we propose a PatchMix data augmentation method to enhance the model’s generalization ability to unseen scenes. Extensive experiments on the KITTI dataset show that the proposed method significantly improves performance over the baseline, particularly in fine-grained scene depth estimation. Moreover, the model also exhibits good generalization performance when transferred to the Make3D and Cityscapes datasets.