EndoRDE: robust self-supervised monocular depth estimation with pseudo guidance for colonoscopic videos

Depth estimation is crucial for the endoscopic scene in navigation, 3D reconstruction, and visualization. Due to the lack of ground truth for many colonoscopic datasets, self-supervised learning is popular in colonoscopic depth estimation tasks. However, colonoscopic data often degrade with low light and less texture compared to common scenarios. To address this, we propose a robust self-supervised depth estimation framework, EndoRDE, with pseudo guidance and mixed data augmentation. The proposed method utilizes the pseudo label generated by the pre-trained foundation model to compensate for the lack of ground truth information. With the pseudo-guide from the foundation model, the bias of the self-supervised framework is amended. In addition, to improve the robustness of the proposed model, the data augmentation branch with an independent depth net is applied to the framework. Two data augmentation methods, frequent spectral map masking and horizontal cut-flip, are applied to the input data of the data augmentation branch. The consistency between two depth nets is then kept to enhance generalization to diverse situations. In experiments on C3VD and SimCol3D datasets, the proposed method obtains state-of-the-art performance on all metrics, with the best MAE and Abs Rel results of 0.022 and 0.077 on the C3VD datasets, and shows its efficiency above other methods.