One-Shot Object Pose Estimation Method for Smart Logistics

Smart logistics is an important branch of intelligent transportation. Recognizing the poses of different goods through vision can help achieve fully automated picking and handling of goods by logistics robots, thereby effectively improving productivity. However, in logistics scenarios, there are often challenges such as diverse types of goods and high annotation costs, which greatly increase the difficulty of the 6-degree-of-freedom pose estimation for the targets. In this paper, we propose a transformer-based one-shot object pose estimation method, which enables fast pose estimation for unknown targets without the need to retrain the network. To obtain sparse 3D point cloud models for targets with few features, we apply the LoftR sparse feature image matching method to the SfM (Structure from Motion) pipeline. We use a similarity network based on attention mechanism to estimate similar poses and further optimize the pose estimation network using the transformer method. Experimental results show that the proposed method outperforms existing one-shot object pose estimation methods in terms of accuracy on the GenMOP, OnePose-Lowtexture data sets, and our custom test data sets. It can be practically applied in smart logistics scenarios.