YOLO-6D-Pose: Enhancing YOLO for Single-Stage Monocular Multi-Object 6D Pose Estimation

Directly regressing 6 degrees of freedom for all the objects from a single RGB image is not well explored. Even end-to-end pose estimation approaches for a single object are inferior compared to state-of-the-art methods in terms of accuracy. Most 6D pose estimation frameworks are multi-stage relying on off-the-shelf deep networks for object and keypoint detection to establish correspondences between 3D object keypoints and 2D image locations. This is followed by applying a variant of a RANSAC-based Perspective-n-Point (PnP) followed by complex refinement operation. In this work, we propose a multi-object 6D pose estimation framework by enhancing the popular YOLOX object detector. The network is end-to-end trainable and detects each object along with its pose from a single RGB image without any additional post-processing. We show that by properly parameterizing the 6D pose and carefully designing the loss function, we can achieve state-of-the-art accuracy without further refinement or any intermediate representations. YOLO-6D-Pose achieves SOTA results on YCBV and LMO dataset, surpassing all existing monocular approaches. We systematically analyze various 6D augmentations to verify their correctness and propose a new translation augmentation for this task. The network does not rely on any correspondences and is independent of the CAD model during inference. Code is available at https://github.com/TexasInstruments/edgeai-yolox.