Lightweight Category-Level 6D Pose Estimation Based on Knowledge Distillation

In realising industrial automation and human-computer interaction, the research on robot grasping operation is fundamental. Existing 6D pose prediction methods have substantially advanced predictive precision, but they often cannot provide robots with sufficiently rich pose information. At the same time, most of these methods rely on complex deep learning models and are difficult to deploy on resource-constrained embedded or edge hardware devices. Most existing 6D pose estimation frameworks output global predictions, such as sparse global features or dense feature representations. However, such global predictions are often difficult to predict by compact student networks accurately. To this end, we propose an innovative knowledge distillation method to extract the global feature distribution of the teacher network into the student network. This not only allows for better training of the student network but also transfers teacher supervision from global predictions to student global predictions, thereby improving the performance of the student network. Compared with the state-of-the-art models, our KDC-Pose occupies merely one-third the storage footprint of the teacher network and demands markedly reduced computational resources. Experiments on multiple benchmarks show that our distillation method can produce better results in a compact student model and performs better in real time.