Position : Home > News > Content

Rui Song’s group and collaborators win the Best Single-model Award in ECCV BOP Challenge

Time: Nov 15, 2022

The 5-day European Conference on Computer Vision (ECCV) was held in Tel Aviv, Israel on Oct. 23. ECCV, along with CVPR and ICCV, is among the three top international conferences in computer vision. The joint team of Xidian University, Ecole Polytechnique Fédérale de Lausanne (EPFL), and Magic Leap won the BOP Challenge (Benchmark for 6D Object Pose Estimation) in the single-model track, andhas been invited to make a presentation in the 6th International Workshop on Recovering 6D Object Pose.


6D object pose estimation

BOP Challenge

The competition attracted teams from many world-renowned universities and research institutions, including Tsinghua University, Queen's University, German Research Center for Artificial Intelligence, Technical University of Munich, Princeton University, Imperial College, and Zhejiang University. Yang Hai,a postgraduate student at Xidian University, won the championship of the track.


The awarded method "Extended FCOS+PFA" adopts a detection-estimation-refinement framework, where the detectordetects all target objects in the input RGB imageand innovatively uses the features of rigid objects to improve the single-stage detector FCOS. It has more advantages over the general detector in 6D pose estimation in the common occlusion case. The estimator estimates the 6D pose of the target object from the target region detected by the detector. This method uses WDR, includes detection pre-processing, and reduces the pose ambiguity present in symmetric objects, resulting in an improvement of approximately 20% compared to the original WDR. The refiner refines the 6D pose by the estimator. This method uses PFAand further includes components such as bi-directional flow and online rendering. It also adopts the idea of iterative refinement, which can achieve a balance between speed and accuracy. In addition, it extends PFA to handle depth data. Compared to the solutions used by other teams, this method is scalable in that it uses only RGB for training and can handle both RGB and RGB-D data in the test time. It significantly outperforms other methods in the single-model track when using either RGB or RGB-D.

Awarded Solution Framework

Qualitative Results

Members of the Team

Yang Hai is a postgraduate student at the Institute of Image Transmission and Processing, school of Telecommunication Engineering, Xidian University, and State Key Laboratory of Integrated Services Networks, under the supervision of Prof. Rui Song. The team is guided by Prof. Rui Song and Associate Prof. Jiaojiao Li from Xidian University, Dr. Mathieu Salzmann and Prof. Pascal Fua from Ecole Polytechnique Fédérale de Lausanne, and Dr. Yinlin Hu from Magic Leap.