Stereo-Assisted Monitored Distillation for Accurate Depth Completion

This paper proposes an innovative approach to depth completion, targeting more accurate depth maps. Our technique enhances the current advanced ensemble distillation method specifically designed for depth completion. We introduce two key components: stereo-assisted monitored distillation and a sparse depth-guided attention module. Central to our approach is the use of a stereo-based model as a “teacher” to improve the accuracy of ensemble distillation, effectively avoiding common errors in completion-based models. Additionally, we integrate a stereo pose estimator to refine structural cues crucial for creating the final distilled depth, a key component in the training process. To further enhance effectiveness, we incorporate an Attention-based Sparse-to-Dense (AS2D) module in the initial layer of the student model, significantly improving the extraction of global features from sparse depth areas. This is particularly beneficial for depth completion accuracy in regions with limited sparse depth points. Extensive experiments validate our method, showing a marked improvement in depth map accuracy over the baseline, affirming the efficacy of our approach in depth completion tasks.