GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation

Gu, Wen; Manhardt, Fabian; Tombari, Federico; Ji, Xiangyang

doi:10.1109/cvpr46437.2021.01634

Cited by 283 publications

(164 citation statements)

References 36 publications

Supporting

Mentioning

164

Contrasting

Order By: Relevance

“…Zhou et.al [74] proposed a continuous 6-dimensional rotation representation that shows advantages over quaternions [44,45] or Lie algebra [20,61] parametrization for neural network training. This representation is utilized in several direct regression works [19,37,67].…”

Section: Rgb-based 6dof Pose Estimationmentioning

confidence: 99%

“…[59,66] used a differentiable depth map and achieved self-supervised network fine-tuning with unlabeled RGB-D data. In an effort to combine correspondence-based methods with direct regression of 6DoF parameters, [67] used correspondence maps as an intermediate geometric representation to regress the pose. [19] further enhances [67] by employing self-occlusion information that provides richer information to predict the object pose with the predicted 2D-3D correspondences.…”

Section: Rgb-based 6dof Pose Estimationmentioning

confidence: 99%

“…To predict the code per pixel in the frame with fine granularity, we process only a Region of Interest (ROI) around object pixels. Following the pipeline similar to [37,41,67], we focus on the object pose and use the available 2D detector predictions to find the ROIs. We crop and resize the ROI from the prediction to a fixed dimension H×W, and apply the exact process to the target vertex code maps during training.…”

Section: Network Architecturementioning

confidence: 99%

“…Now we incorporate the predicted code and visible mask and the reference 3D model encoding to match correspondences. Different from common dense correspondences such as [49,67,68], this compact representation also enables a bijective correspondence between the surface vertices and the descriptor space. That means, unlike the regressed 3D point which can be off the object surface, our estimated 3D correspondences always refers to a vertex on the object model, which eases the matching stage for the pose solver.…”

Section: Pose Estimationmentioning

confidence: 99%

“…As the physically-based rendering (pbr) training images [18] for both datasets are publicly accessible now, using pbr images to support the training can help us focus on the pose estimation CNN itself. We use the pbr images together with the real images for the training in the same manner as [19,29,67].…”

Section: Experiments Setupmentioning

confidence: 99%

See 4 more Smart Citations

ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation

Su¹,

Saleh²,

Fetzer³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Establishing correspondences from image to 3D has been a key task of 6DoF object pose estimation for a long time. To predict pose more accurately, deeply learned dense maps replaced sparse templates. Dense methods also improved pose estimation in the presence of occlusion. More recently researchers have shown improvements by learning object fragments as segmentation. In this work, we present a discrete descriptor, which can represent the object surface densely. By incorporating a hierarchical binary grouping, we can encode the object surface very efficiently. Moreover, we propose a coarse to fine training strategy, which enables fine-grained correspondence prediction. Finally, by matching predicted codes with object surface and using a PnP solver, we estimate the 6DoF pose. Results on the public LM-O and YCB-V datasets show major improvement over the state of the art w.r.t. ADD(-S) metric, even surpassing RGB-D based methods in some cases.

show abstract