6-D Pose Detection Based on Bidirectional Fusion of Texture and Depth Information

In dynamic and complex real-world scenarios, RGB-D based object (6D) pose estimation has garnered substantial attention for its wide-ranging applications and robust performance. Keypoint-based methods, in particular, have been favored for their superior accuracy. However, these methods often falter when dealing with heavily occluded or truncated target objects. To overcome these limitations, we propose a novel bimodal fusion network architecture. Our method begins with object detection to precisely locate the target object, followed by random sampling within the detected region. A dedicated feature extraction network then separately extracts visual features from RGB images and geometric features from depth images. These features are subsequently integrated through a bimodal fusion mechanism, enabling accurate pose prediction via keypoint estimation. Through extensive experiments on the LineMOD and Occlusion LineMOD datasets,our method has demonstrated significant improvements in pose estimation accuracy. These results confirm the efficacy of the bimodal feature fusion strategy in enhancing keypoint-based 6D pose estimation using RGB-D data.