Haonan Dong 董皓男

I will be an engineer at XGRIDS in July, where I will continue my work in 3D computer vision. Before that, I got my master degree at CVRS(Computer Vision & Remote Sensing Lab), Wuhan University, where I work on 3D computer vision and I am supervised by Prof. Jian Yao. In July, 2022, I was fortunately worked with Prof. Lingjie Liu on NeRF.

I am interested in most 3D techs including Neural Radiance Field (NeRF), Multi-View Stereo(MVS), Structure from Motion (SfM) and Augmented Reality (AR). My dream is to build an automatic machine for measuring all the 3D data in our daily life.

I also finished my bachelor's degree in Wuhan University. Awards and more information are listed in my CV.

Email  /  CV  /  Github

I'm interested in 3D computer vision. Much of my research is about inferring the physical world (structure, motion, shape, etc) from images by Neural Radiance Field (NeRF), Structure from Motion (SfM) and Multi-view Stereo (MVS). Representative papers are highlighted.

We present a circular-pattern calibration board that has rotational invariance. Experiments prove that this board with the proposed pipeline is all you need for the comprehensive optical camera calibration task.

An approach for view-graph key-subset extraction is proposed to address the data redundancy and uneven distribution problem in SfM. Experiments conducted on a variety of datasets demonstrate the advantages of our approach in efficiency, robustness and even accuracy compared to SOTA methods.

We present several robust unsupervised learning loss functions for Multi-View Stereo and our method reaches the performance of the state-of-the-art.

Our algorithm can preferentially evacuate the rooms near the fire origin，avoid the indoor congestion，and ensure the stability of the evacuation plan.

I work with my teammates to do some interesting things with 3D technologies. Representative works are highlighted.

By integrating NeuS into the instant-ngp framework, and the performance is 6X faster than the official implementation. Currently I work on the density-grid accelerating by the ray marching algorithm to further enhance the performance..

We propose a novel strategy to optimize the depth computed from a RGB-D camera and the low-accuracy poses from a lightweight online VO. After the optimization, the poses and the depth are precise enough to rebuild a high-quality model with a simple RGB-D hardware, say the iPad Pro with the TOF camera.

Currently, we are designing a cheap calibration solution for the multi-camera intrinsic and extrinsic calibration, from the pinhole to the fish-eye, from the monocular to the stereo..

We present a whole pipeline to get real model (eg. depth maps, textured meshes, etc.) from multi-view images or videos, and other applications including getting the real-scale maps for the AR usage.

Using the printed ArUco Marker, we optimize the whole sparse point cloud by bundle adjustment with the real scale. Experiments show that our method can recover the real scale of the map and can be directly used in the AR applications..

Paper Reviewer : IROS 2022, ISPRS

RGB 3D Reconstruction Pipeline : Currently, I developed a software based on classic open-src codes to reconstruct a textured mesh from video or RGB image sequences.

Depth Fusion : This software can fuse the depth from multi-view stereo result or sensors with known poses with two strategies. One is to use the view selection strategy of COLMAP, and the other is to merge the neighboring two images.

Awesome Indoor Reconstruction : A curated list of awesome indoor reconstruction methods and datasets.