Multi Task-Guided 6D Object Pose Estimation

Object pose estimation remains a fundamental challenge in computer vision, with cutting-edge methods relying on both RGB and depth data. Depth information is pivotal, offering crucial geometric cues that enable algorithms to navigate occlusions, fostering a more comprehensive scene under-standing and precise pose estimation. However, RGBD-based methods often require specialized depth sensors, which can be costlier and less accessible compared to standard RGB cameras. Consequently, research has explored techniques aiming to estimate object pose solely from color images. Yet, the absence of depth cues poses challenges in handling occlusions, comprehending object geometry, and resolving ambiguities arising from similar colors or textures. This paper introduces a end-to-end multi-task-guided object pose estimation method, utilizing RGB images as input and producing the 6D pose of multiple object instances. While our approach employs both depth and color images during training, inference relies solely on color images. We incorporate depth images to supervise a depth estimation branch, generating depth-aware features further refined through a cross-task attention module. These enhanced features are pivotal for our object pose estimation. Our method's innovation lies in significantly enhancing feature discriminability and robustness for object pose estimation. Through extensive experiments, we demonstrate competitive performance compared to state-of-the-art methods in object pose estimation.