Multi-Scale Distillation for Low Scanline Resolution Depth Completion

Depth information is crucial for an autonomous driving system as it helps the system understand the environment and make decisions. Most deep learning-based depth completion methods are primarily designed for high-resolution lidars (e.g. 64 scanlines). However, when the number of lidar scanlines decreases, such as with 32 scanlines or 16 scanlines lidars, existing solutions may face challenges in reliably predicting dense depth maps. To address this issue, this paper proposes an effective framework based on knowledge distillation, which incorporates mixed-scanline resolution training and feature-level fusion to train a powerful teacher network that dynamically fuses features from high-scanline resolution and low-scanline resolution inputs. By supervising the student network based on the guidance of the teacher network, the knowledge from the multi-scale fusion teacher network is effectively transferred to the low-scanline resolution student network. For the inference process, only the student network is utilized. The proposed framework has been applied to various existing depth completion networks. The experimental results show the effectiveness of the proposed method by using the KITTI dataset, which shows that it can serve as a universal framework for depth completion tasks.