UDCGN: Uncertainty-Driven Cross-Guided Network for Depth Completion of Transparent Objects

In the field of robotics, most perception methods rely on depth information captured by RGB-D cameras. However, the ability of depth sensors to capture depth information is hindered by the reflection and refraction of light on transparent objects. Existing methods of completing transparent objects' depth information are usually impractical due to the need for fixtures or unacceptably slow inference speeds. To address this challenge, we propose an efficient multi-stage architecture called UDCGN. This method progressively learns completion functions from sparse inputs by dividing the overall recovery process into more manageable steps. To enhance the interaction between different branches, Cross-Guided Fusion Block (CGFB) is introduced into each stage. The CGFB dynamically generates convolution kernel parameters from guided features and convolutes them with input features. Furthermore, the Adaptive Uncertainty-Driven Loss Function (AUDL) is developed to handle the uncertainty issue of sparse depth. It optimizes pixels with high uncertainty by adapting different distributions. Comprehensive experiments on representative datasets demonstrate that UDCGN significantly outperforms state-of-the-art methods in terms of both performance and efficiency.