Recently,“EfficientPose: Efficient Human Pose Estimation with Neural Architecture Search”co-authored by EIC graduate student Zhang Wenqiang and doctoral student Fang Jiemin won Computational Visual Media (CVMJ) Best Paper Award. The corresponding authors are Associate Professor Wang Xinggang and Professor Liu Wenyu from EIC.

The paper focuses on human pose estimation in video images, an important research topic in the field of visual media, which can be extensively applied in metaverse, virtual reality, game interaction and life support. The paper proposes EfficientPose, a new efficient computing framework. Compared with the traditional pose estimation methods, this framework has significantly improved in computing efficiency and maintained high prediction accuracy at the same time. EfficientPose mainly includes two parts: efficient backbone network and efficient head network. Through differential network architecture search, the backbone network is designed automatically at a very low computing cost. In the optimization process, the backbone architecture can not only make specific adjustment for the pose estimation task, but also optimize the computing efficiency of the architecture; In addition, the paper also proposes a spatial information correction module, which effectively solves the chessboard effect in prediction at a low cost, so as to improve the prediction accuracy.

CVMJ is a new domestic journal jointly published by Tsinghua University and Springer. The journal publishes original and high-quality research papers and review papers in the field of visual media, covering various topics such as computer vision, computer graphics, image/ video processing and machine learning. The selection of CVMJ Best Paper is strict: the associate editors first nominate 11 papers out of all the published papers in 2021, and experts invited by the editor select the only one paper as the best paper of 2021.

More information about CVMJ: http://cvm.tsinghuajournals.com/EN/column/column5214.shtml