Compact Depth-wise Separable Precise Network For Depth Completion

Predicting a dense depth map from synchronized LiDAR scans and RGB images using compact deep neural networks presents significant challenge. While most state-of-the-art models enhance prediction accuracy by increasing the number of parameters, leading to substantial memory consumption, completion tasks in areas such as autonomous driving primarily utilize edge devices powered embedded GPUs. In this paper, we introduce methodology for creating an efficient, high-fidelity model derived base model. Our proposed replaces conventional convolutional encoder layers with depth-wise separable convolutions, transposed decoders up-sampling plus convolution. We further employ random layer pruning stability test, guiding design our architecture preventing over-parameterization. Additionally, straightforward yet robust knowledge distillation method network performance improve scalability meet higher quality requirements. experimental results demonstrate improvement over existing terms performance, while significantly reducing parameters compared larger models.