ACF-Net: Attention Context Fusion Network for 6D Pose Estimation

The advent of cost-effective RGB-D sensors has made it more feasible to achieve high-precision 6D pose prediction from RGB-D images. One of the primary research challenges entails extracting distinctive features from the color features presented in RGB images, as well as capturing geometric details from depth data, while effectively harnessing the synergistic potential of these two mutually complementary data sources. In this study, we propose a bidirectional attention context fusion framework for learning 6D pose parameters from RGB-D images. Specifically, we introduce a lightweight attention mechanism to improve the ability of appearance features learning from RGB images. Additionally, we design a features extraction framework of point clouds to fully explore the spatial information of the local distance structure within the neighborhood and more representative global features obtained by hybrid contextual local information. Furthermore, we combine the speed-up robust feature algorithm with the farthest point sampling algorithm, taking full advantage of our appearance and geometric information extraction to predict pose estimation. In the experimental results, we provide compelling evidence that our method surpasses the performance of state-of-the-art methods on two demanding benchmark datasets, LineMOD and YCB-Video.