Depth Completion of Transparent Objects Based on Feature Fusion

Accurately perceiving the depth of transparent objects is crucial for modern automation and robotics but is challenging due to their optical properties. To address this challenge, This paper proposes an MFAU-Net-based depth completion method for transparent objects, utilizing geometric features from RGB images and introducing the Encoder Attention Module (EAM) and Feature Fusion Module (FFM) for effective multi-level feature extraction and fusion. Compared to the ClearGrasp model, the RMSE improves by 15.97% and the MAE by 14.81%. Experimental results show that MFAU-Net significantly enhances depth estimation precision and resilience, providing a novel approach for transparent object depth perception.