A high-resolution network-based approach for 6D pose estimation of industrial parts

Abstract

The estimation of 6D pose of industrial parts is a fundamental problem in smart manufacturing. Traditional approaches mainly focus on matching corresponding key point pairs between observed 2D images and 3D object models via hand-crafted feature descriptors. However, key points are hard to discover from images when the parts are piled up in disorder or occluded by other distractors, e.g., human hands. Although the emerging deep learning-based methods are capable of inferring the poses of occluded parts, the accuracy is not satisfactory largely due to the loss of spatial resolution from multiple downsampling operations inside convolutional neural networks. To overcome this challenge, this paper proposes a 6D pose estimation model consisting of a pose estimator and a pose refiner, by leveraging High-Resolution Networks as the backbone. Experiments are further conducted on a dataset of industrial parts to demonstrate its effectiveness.