Figure 1: Results display: input images in the first column, predicted 3D bounding boxes in the second column, and 3D scene reconstruction in the third column.

The finalist paper, Total3DUnderstanding: Joint Layout, Object Pose and Mesh Reconstruction for Indoor Scenes from a Single Image, was prepared by CUHK-Shenzhen scholars in collaboration with those from the Shenzhen Research Institute of Big Data (SRIBD), Bournemouth University, and Xiamen University. The first author is Yinyu Nie, a visiting student from Xiamen University, with Dr. Xiaoguang Han as the sole corresponding author.

The paper proposes an end-to-end solution to jointly reconstruct room layout, object bounding boxes and meshes from a single image. The method builds upon a holistic scene context and proposes a coarse-to-fine hierarchy with three components: 1. room layout with camera pose; 2. 3D object bounding boxes; 3. object meshes. It is demonstrated that this method consistently outperforms existing methods in all the components. For this reason, the paper was highly appreciated by the reviewers, and accepted as an Oral Paper and ultimately a Paper Award Nominee by CVPR2020.

Details of the paper can be found at:

https://arxiv.org/abs/2002.12212

For a description in Chinese, please see: Oral:：CVPR'20 Oral: 一张照片三维重建你的房间 | 将门好声音

Another work by Dr. Xiaoguang Han's team, FPConv: Learning Local Flattening for Point Convolution, was also accepted for inclusion in CVPR2020. This study was conducted by CUHK-Shenzhen in collaboration with the Shenzhen Research Institute of Big Data (SRIBD), the University of Science and Technology of China, and Kuaishou Technology, with Prof. Shuguang Cui and Dr. Xiaoguang Han as co-directors. The first author, Yiqun Lin, is a graduate student (Mphil) at the University, and Dr. Xiaoguang Han serves as the corresponding author.

Details of the paper can be found at:

https://arxiv.org/pdf/2002.10701.pdf

For a description in Chinese, please see: ：三维变二维，无需体素化！港中文提出用于点云卷积的局域展平网络模块FPConv

The research work shortlisted for Best Paper Finalist by CVPR2019 is A Skeleton-bridged Deep Learning Approach for Generating Meshes of Complex Topologies from Single RGB Images.

The paper was co-authored by scholars from CUHK-Shenzhen, Shenzhen Research Institute of Big Data, South China University of Technology and Microsoft Research Asia, and the first author goes to Jiapeng Tang, a summer visiting student, with Doctor Xiaoguang Han as the co-first author. The paper proposes a skeleton-bridged, stage-wise learning approach to address the challenging task of learning 3D object surface reconstructions from single RGB images. The work was well received by the conference reviewers, with all three reviewers giving a Strong Accept opinion.

Details of the paper can be found at:

https://arxiv.org/pdf/1903.04704.pdf

For a description in Chinese, please see：CVPR 2019 | 基于骨架表达的单张图片三维物体重建方法

About the Professor

Prof. Xiaoguang HAN

Research Assistant Professor of SSE

Dr. Xiaoguang Han, a research assistant professor of the Chinese University of Hong Kong, Shenzhen and a research scientist in the  Shenzhen Research Institute of Big Data (SRIBD). His research interests include computer vision, computer graphics, virtual reality and medical image processing, etc., in the direction of the famous international journals and conference papers published nearly 30, including top-level meetings and journal SIGGRAPH, CVPR, ICCV, AAAI, ACM TOG, IEEE TIP, IEEE TVCG, etc. His work won the Best Demo of emerging technologies 2013 of the top computer graphics conference Siggraph Asia , selected as one of the best computing papers of the year 2016 and was nominated as the best paper for two consecutive years at the top computer vision conference CVPR in 2019 and 2020 (The turnout was 0.8% and 0.4%, respectively), and his team won the IEEE ICDM Global Weather Challenge in November 2018 (More than 1,700 teams participated).