Multi-view Improved Monitored Distillation for Depth Completion

This paper proposes a new depth completion method based on multi-view improved monitored distillation to generate more accurate depth maps. Based on the state-of-the-art depth completion method named ensemble distillation, we introduce an existing stereo-based model as a teacher model to improve ensemble distillation accuracy and generate a more accurate student model in training by avoiding inherent error modes of completion-based teachers as well as minimizing the reconstruction error for a given image. We also leverage multi-view depth consistency and multi-scale minimum reprojection to provide self-supervised information. These methods use the existing structure constraints to yield supervised signals for student model training without great expense on gathering ground truth information of depth. Our extensive experimental evaluation demonstrates that our proposed method can effectively improve the accuracy of baseline method of monitored distillation.