Wanshui Gan

I am a 3rd year PhD student at the Sugiyama-Yokoya-Ishida Lab at the University of Tokyo, Department of Complexity Science and Engineering, advised by Prof. Naoto YOKOYA. I am also the Junior Research Associate of the Geoinformatics Team at the RIKEN Center for Advanced Intelligence Project (AIP).

Prior to that, I received the B.S. degree from the Guangdong University of Technology, China, in 2018 and the M.S. degree from the University of Macau, China, in 2021. My previous research interest lies in 3D vision, large scene parsing, and reconstruction. Additionally, I am interested in exploring 3D foundation models and 3D generative models. You are welcomed to contact me by email if you are interested in my work or potential collaboration.

Email  /  Google Scholar  /  Github  /  Twitter

* indicates equal contribution

We introduce GaussianOcc, a systematic method that investigates the two usages of Gaussian splatting for fully self-supervised and efficient 3D occupancy estimation in surround views. The proposed GaussianOcc method enables fully self-supervised (no ground truth pose) 3D occupancy estimation in competitive performance with low computational cost (2.7 times faster in training and 5 times faster in rendering).

We introduce a comprehensive framework for surrounding-view 3D occupancy estimation, 3D reconstruction and depth estimation via volume rendering, featuring network design, loss design, and evaluation metric based on discrete point level sampling.

We propose the method V4D, a simple yet effective and efficient framework, for 4D novel view synthesis with the 3D voxel, which directly models the 4D neural radiance field without the need for canonical space.

We introduce a novel 6D pose estimation framework, ES6D, based on the XYZNet and A(M)GPD loss. The XYZNet is designed for feature extraction from RGB-D data. It has a fully convolutional architecture and achieves an excellent trade-off between efficiency and effectiveness. Additionally, the A(M)GPD loss is proposed to handle symmetric objects, and performs better than ADD(S) loss.

We propose a novel light-weight adaptive network (LWANet) for real-time stereo depth estimation, achieving competitive performance compared with MADNet and StereoNet, and it has the advantages of low computational cost and low GPU memory space.

Website Template