Hongbin Xu

I am a second year PHD student in the Department of Automation in South China University of Technology (SCUT), advised by Prof. Wenxiong Kang

I am broadly interested in computer vision and deep learning. My current research focuses on:

3D reconstruction: especially for Multi-view Stereo (MVS) (Explicit 3D reconstruction) and Neural Radiance Field (NeRF) (Implicit 3D reconstruction).

Self-supervised Learning: especially for unsupervised/self-supervised approaches for tasks like depth estimation (monocula/stereo/multi-view).

3D Vision: especially for 3D representation learning for vision applications such as 3D biometrics, point cloud perceptions, and etc.

Email / Google Scholar / Github

Experience

2020-2021: Research Intern in Multimedia Laboratory (MMLAB) of Shenzhen Institutes of Advanced Technology (SIAT), University of Chinese Academy of Sciences (UCAS), advised by Prof. Yu Qiao .

2021-2022: Research Intern in Alibaba Damo Academy, advised by Zhipeng Zhou .

Preprint

* indicates equal contribution

Semi-supervised Deep Multi-view Stereo
Hongbin Xu, Zhipeng Zhou, Weitao Chen, Baigui Sun, Yang Liu, Hao Li, Wenxiong Kang

[Arxiv] [Code] [Data/Benchmark]

We firstly explore a novel semi-supervised setting of learning-based MVS problem that only a tiny part of the MVS data is attached with dense depth ground truth. The semi-supervised MVS (Semi-MVS) problem may break the basic assumption in semi-supervised learning that unlabeled data and labeled data share the same label space and data distribution. To handle these issues, we propose a novel semi-supervised MVS framework, namely SE-MVS. We evaluate our methods on existing benchmarks (DTU, Tanks&Temples) and self-built MVS 3D reconstruction benchmarks (BlendedMVS, GTASFM).

Publications

* indicates equal contribution

Self-supervised Multi-view Stereo via Effective Co-Segmentation and Data-Augmentation
Hongbin Xu*, Zhipeng Zhou*, Yu Qiao, Wenxiong Kang, Qiuxia Wu
Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021, Distinguished Paper Award
[Paper] [Arxiv] [Code]

We propose a self-supervised/unsupervised framework for learning-based Multi-view Stereo (MVS). To handle the color constancy ambiguity problem in previous self-supervised MVS methods, we propose a novel framework integrated with more reliable supervision guided by semantic co-segmentation and data-augmentation. Experimental results show that our proposed method can achieve competetive performance compared with state-of-the-art fully-supervised MVS methods as well as the unsupervised methods.

Digging into Uncertainty in Self-supervised Multi-view Stereo
Hongbin Xu, Zhipeng Zhou, Yali Wang, Wenxiong Kang, Baigui Sun, Hao Li, Yu Qiao
IEEE International Conference on Computer Vision (ICCV), 2021
[Paper] [Arxiv] [Code]

We propose to estimate epistemic uncertainty in self-supervised MVS, accounting for what the model ignores. Considering the two problems: ambiguious supervision in foreground and invalid supervision in background, we propose an uncertainty-guided self-supervised MVS framework, namely U-MVS. Extensive experiments show that our U-MVS framework achieves the best performance among state-of-the-art unsupervised methods.

Website Template