CKDNet: A Cascaded Knowledge Distillation Framework for Lightweight Stereo Matching

In this paper, we propose an efficient cascaded knowledge distillation framework which is tailored for the refinement of lightweight stereo matching network. Lightweight network is the feasible means to achieve real-time depth estimation on embedded devices, but often suffered from reduced accuracy. Despite previous attempts to enhance performance through the introduction of new modules, the resulting improvements are modest. To address this challenge, our framework integrates knowledge distillation with three models to cascaded optimize the performance of lightweight architectures. Our contributions include three folds: First, we demonstrate the effectiveness of knowledge distillation for stereo matching and its transferability across different scale models. Second, we conduct a comprehensive comparison of distillation methods and propose the cascade distillation framework to improve the accuracy of lightweight model. Third, we proposed an adjustable compression module to enable flexible changes of model size by demands. Through extensive experimentation and evaluation, we validate the efficacy of our approach. Our lightweight model has archived 0.87px on Sceneflow(EPE), and 2.26% on KITTI2015(D1). And only 12 ms for a single-time inference.