An Improved Transparent Object Pose Estimation Method for Materialization Laboratory Applications

To mitigate the degradation of pose estimation accuracy caused by the optical characteristics of transparent instruments in materialization laboratories, an instance-level pose estimation algorithm based on multi-stage and multi-scale texture feature extraction is proposed. The framework integrates HRNet and ASPP modules to extract fine-grained texture features from RGB images, while pose prediction is refined through keypoint detection and multi-stage feature fusion. Final 6D poses are computed using the PnP algorithm with known 3D models. The proposed method is trained and evaluated on a synthetic transparent object dataset generated via NVISII and further validated in real-world experiments using a dual-arm robotic platform. Experimental results demonstrate superior accuracy and generalization capability in complex scenarios, confirming the method's effectiveness and robustness for intelligent robotic operations in materialization laboratory environments.