关于我

Xiao Liu (刘骁)

E-mail: liuxiao (at) foxmail.com

I joined the PICO department at ByteDance in December 2021 as the head of the Mixed Reality (MR) algorithms team. Prior to that, I spent four years (from October 2017 to December 2021) at Megvii (Face++) Research, serving as the Director responsible for 3D vision algorithm research. Between December 2010 and October 2017, I worked at Tencent Technology (Beijing) Co., Ltd., contributing to both the WeChat and Tencent Research departments as a pattern recognition algorithm researcher.

I earned my Master's degree in Computer Science from BeiHang University in January 2011, following my Bachelor's degree in Science from the same institution in July 2007.

My research interests span 3D vision (encompassing 3D reconstruction, 3D AIGC, and 3D scene understanding), computer vision, SLAM (Simultaneous Localization and Mapping), robotics, among others.

Interests

Computer Vision, 3D Vision, SLAM, AR/VR, Robotics

Professional Experience

时间公司部门职位
2021.12-至今字节跳动PICOMR 算法研究员
2017.10-2021.12旷视科技有限公司研究院3D 算法研究员
2013.10-2017.10腾讯科技（北京）有限公司微信产品部高级研究员
2010.12-2013.10腾讯科技（北京）有限公司研究院高级研究员
2010.06-2010.11佳能信息（北京）有限公司研究院OCR 研究员（实习）
2009.09-2010.04联想研究院移动互联网技术部研究员（实习）

Education

时间学校院系学历
2008.9-2011.1北京航空航天大学计算机科学与技术硕士研究生
2003.9-2007.7北京航空航天大学应用物理（应用电子技术）本科

Research

Challenge Winnings

年份比赛链接
2024CVPR 2024 3D Open-Vocabulary Scene Understanding (OpenSUN3D) Challenge
Zhishan Zhou, Yunke Cai, Chunjie Wang, Xiaosheng Yan, Bingwen Wang, Xiao Liu
PICO-MR Team
1st place in 3D Functionality Grounding
2nd place in 3D Object Instance Search		Workshop | Challenge Track 1 Track 2 | More >>
2024CVPR 2024 Monocular Depth Estimation Challenge
GuangYuan Zhou, ZhengXin Li, Qiang Rao, YiPing Bao, Xiao Liu
1st place		Results | Workshop | Challenge | More >>
2023ICCV 2023 3D Open-Vocabulary Scene Understanding (OpenSUN3D) Challenge
Hongbo Tian, Chunjie Wang, Xiaosheng Yan, Bingwen Wang, Xuanyang Zhang, Xiao Liu
1st place		Results | Workshop | Challenge | More >>
2021ICCV 2021 Long-Term Visual Localization under Changing Conditions Challenge
Shuxue Peng, Zihang He, Haotian Zhang, Ran Yan, Chuting Wang, Qingtian Zhu, Yikang Ding, Liangtao Zheng and Xiao Liu
1st place in ​Outdoor visual localization​ track
1st place in Indoor visual localization track		Website | Workshop | Recording | More >>
2021ICCV 2021 Map-Based Localization for Autonomous Driving Challenge
Megvii 3D Team
1st place		Website | Workshop | Recording | More >>
2021CVPR 2021 Image Matching Challenge
Xiaopeng Bi, Yu Chen, Xinyang Liu, Dehao Zhang, Ran Yan, Zheng Chai, Haotian Zhang, and Xiao Liu
1st place in Unlimited keypoints category
1st place in SimLocMatch competition
2nd place in Restricted keypoints category		Website | Recording
1. IMC 2021 Challenge | Report
2. SimLocMatch Challenge | Report
2021IROS 2021 THE HILTI SLAM Challenge
Xiaojia Xie, Xizhen Xiao, Xuefeng Shen, Wei Zuo, and Xiao Liu
1st place		Website | Workshop | Report | More >>
2020CVPR 2020 SLAM Challenge
Haotian Zhang, Zheng Chai, and Xiao Liu
1st place in Mono Track competition
1st place in Stereo Track competition		Website | Recording | More >>

Recent Publications

论文/作者会议/期刊下载
Coin3D: Controllable and Interactive 3D Assets Generation with Proxy-Guided Conditioning
Wenqi Dong, Bangbang Yang, Lin Ma, Xiao Liu, Liyuan Cui, Hujun Bao, Yuewen Ma, Zhaopeng Cui		SIGGRAPH 2024Paper
Multi-modal Relation Distillation for Unified 3D Representation Learning
Huiqun Wang, Yiping Bao, Panwang Pan, Zeming Li, Xiao Liu, Ruijie Yang, Di Huang		ECCV 2024Paper
DreamSpace: Dreaming Your Room Space with Text-Driven Panoramic Texture Propagation
Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, Yuewen Ma		IEEE VR 2024Paper
Tri-MipRF: Tri-Mip Representation for Efficient Anti-Aliasing Neural Radiance Fields
Wenbo Hu, Yuling Wang, Lin Ma, Bangbang Yang, Lin Gao, Xiao Liu, Yuewen Ma		ICCV 2023 Oral
Bestpaper Finalist 17/8260		Paper
MegBA: A High-Performance and Distributed Library for Large-Scale Bundle Adjustment
Jie Ren, Wenteng Liang, Ran Yan, Luo Mai, Shiwen Liu, Xiao Liu		ECCV 2022Paper
TransMVSNet: Global Context-aware Multi-view Stereo Network with Transformers
Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, Xiao Liu		CVPR 2022Paper
ELSD: Efficient Line Segment Detector and Descriptor
Haotian Zhang, Yicheng Luo, Fangbo Qin, Yijia He, Xiao Liu		ICCV 2021Paper
M^3VSNet: Unsupervised Multi-metric Multi-view Stereo Network
Baichuan Huang, Hongwei Yi, Can Huang, Yijia He, Jingbin Liu, Xiao Liu		ICIP 2021Paper
Structure Reconstruction Using Ray-Point-Ray Features:
Representation, Triangulation and Camera Pose Estimation
Yijia He, Xiangyue Liu, Xiao Liu, Ji Zhao		ICRA 2021
Monocular Visual SLAM with Points and Lines for Ground Robots in Particular Scenes: Parameterization for Lines on Ground
Meixiang Quan, Songhao Piao, Yijia He, Xiao Liu & Muhammad Zuhair Qadir		J Intell Robot Syst
101, 72 (2021)
TP-LSD: Tri-Points Based Line Segment Detector
Siyu Huang, Fangbo Qin, Pengfei Xiong, Ning Ding, Yijia He, Xiao Liu		ECCV 2020Paper
Leveraging Planar Regularities for Point Line Visual-Inertial Odometry
Xin Li, Yijia He, Jinlong Lin, Xiao Liu		IROS 2020Paper
MLIFeat: Multi-level information fusion based deep local features
Yuyang Zhang，Jinge Wang，Shibiao Xu，Xiao Liu，Xiaopeng Zhang		ACCV 2020Paper
Coarse-To-Fine Visual Localization Using Semantic Compact Map
Ziwei Liao, Jieqi Shi , Xianyu Qi, Xiaoyu Zhang, Wei Wang, Yijia He, Xiao Liu , Ran Wei		ICCR 2020Paper

Talks

TalkTalk地址
将门公开课 2020三维视觉与机器人链接
SIGGRAPH Asia 2016QAR Open Platform

链接

    24 Comments
    1. yzq

      【请教】有关你写的MXNet的image-classification-predict.cc的疑问

      liuxiao你好，

      我看到你贡献给MXNet的image-classification-predict.cc代码做predict。

      https://github.com/dmlc/mxnet/blob/master/example/cpp/image-classification/image-classification-predict.cc

      有个疑问请教一下。

      在GetMeanFile中你是直接设置mean=117，你的注释里也写着：

      // Better to be read from a mean.nb file

      我现在有图像均值的二进制文件mean.bin，想直接整张图片做矩阵减，

      还请问该如何读文件 以及做后续的norm操作呢？

      直接用你的数值在我的数据集上预测结果不佳，
      而且实验发现修改该值对最终的top-1预测结果影响很大

      希望能够获得你的帮助，非常感谢！

      回复
    2. WeiSheng

      Hi 技术刘, skylook,

      你们好! 请问能帮忙看看以下的问题吗?
      https://github.com/dmlc/mxnet/issues/2709
      谢谢!
      WeiSheng

      回复
    3. jacob

      您好我有个关于image-classification-predict.cc的疑问
      我想在windows环境下运行这个demo但是在windows下编译的dll跟lib使用时会报error LNK2001: 无法解析的外部符号 __imp__MXPredFree D:\poi\testcv_mxnetnet\testcv\Source.obj
      这样错误，用官方预编译的版本也会这样，怎么解决啊？

      回复
      • skylook

        抱歉，没有用 windows 编译过 mxnet。

        回复
    4. 许先生

      liuxiao.net

      回复
    5. chen

      你好，这篇文章里，你对比了vi-orb-slam和orb slam的定位精度，但是我理解orb slam应该是得不到尺度信息的，而vi-orb-slam是有尺度的，你给出的rmse应该是实际的物理值，那orb slam的rmse值是怎么得到的呢？ 另外，rmse是指的关键帧的计算位置和ground truth的位置的误差吗？ 谢谢

      回复
    6. chen

      liuxiao 你好，你的这篇文章里，你对比了vi-orb-slam和orb slam的定位精度，但是我理解orb slam应该是得不到尺度信息的，而vi-orb-slam是有尺度的，你给出的rmse应该是实际的物理值，那orb slam的rmse值是怎么得到的呢？ 另外，rmse是指的关键帧的计算位置和ground truth的位置的误差吗？ 谢谢

      回复
      • skylook

        orb slam双目可以估计绝对尺度，单目是用拟合算法拟合出一个与真值最接近的scale作为绝对尺度再进行估计。这些方法在前面的博文中evo工具均有提供。rmse是关键帧的位置和GT的误差

        回复
    7. xingbo WANG

      1. orb-slam2 保存的相机轨迹和groundtruth有很大的区别，完全不吻合。是不是坐标系的问题啊？orb-slam2没做坐标变换？
      2. orb-slam2 有木有加入IMU信息的版本啊？

      回复
      • skylook

        1、如果运行正确没有问题，请参见：https://www.liuxiao.org/2017/11/%E4%BD%BF%E7%94%A8-evo-%E5%B7%A5%E5%85%B7%E8%AF%84%E6%B5%8B-vi-orb-slam2-%E5%9C%A8-euroc-%E4%B8%8A%E7%9A%84%E7%BB%93%E6%9E%9C/
        2、加入IMU信息的没有官方版本只有王京等一些人个人的一些实验代码。

        回复
    8. lianglong

      hello，关于mxnet里的example。我想在自己的系统里面使用mxnet的结果，做相近的sementation的工作，只需要将lib文件里编译好的库文件复制到本地文件就可以了吗？那cmakelist里面需要怎么链接呢？看您提供的makefile不是很明白。

      回复
    9. 许先生

      liuxiao.net出售

      回复
    10. 匿名

      有没有C风格的调用mxnet的例子, 想用rust 绑定一下
      (嘤嘤怪出没)

      回复
      • skylook

        最近许久没有再研究mxnet了

        回复
    11. 匿名

      单目是用拟合算法拟合出一个与真值最接近的scale作为绝对尺度再进行估计。大佬，这句话我没有太理解，有没有参考的资料，我学习下。最近有点纠结这个单目的scale。谢谢

      回复
    12. 刘敏

      Hi 刘先生

      我是无人驾驶领域猎头刘敏，关注到您在腾讯的工作，我正在帮我北上深杭苏的无人驾驶客户寻找SLAM这块的职位，期望能与您交换个电话号码做进一步沟通，我的电话是15959217633（微信同号），期待您的消息。

      刘敏

      回复
      • skylook

        哇，现在猎头同志这么专业都上博客了

        回复
        • 匿名

          666 tql

          回复
        • 十三散人

          我也是，哈哈

          回复
    13. jbin

      你好，巧合下看到你的博客，想了解一下在deep vslam这块的一些问题。UnSuperPoint在我看到的几个复现中都表示学习不出特征的位置，结果是一个均匀分布的检测结果。所以想问一下博主关于UnSuperPoint里面，点对选取和score map的选择。看了解读的文章还是不太明确，像GCN的思路就是直接用其它的关键点检测方法，得到原图的特征点，然后进行投影得到变换图的数据。这个思路应该是好理解的。但是UnSuperPoint或者SuperPoint似乎没有用第三方的特征点检测方法，那如何确定关键点的位置和描述子呢。

      回复
      • skylook

        SuperPoint 是通过在仿真数据集上训练一个可以预测几何角点的 MagicPoint 再用它在真实数据集做 Homography Adaptation 获得真值，本质上是一个弱标记的方式。具体在他的论文中讲得还是很清楚的。
        UnSuperPoint 是完全的无监督方式，具体参见 Figure 3，它构造的 Loss 本质上是学习在各种不同 Homography 变化下依然能够匹配的点，这种方式在其他文章中也出现过。对于后者由于没有实现过，是否真的可行以及是否实际上需要预训练参数，还是存疑的。

        回复
    14. 匿名

      搞技术的好牛逼啊，真是佩服，北航师弟膜拜

      回复

    Add a Comment

    您的邮箱地址不会被公开。 必填项已用 * 标注