Benjin Zhu (朱本金)

I’m a researcher at the BaseDetection Team of MEGVII Technology, focusing on Computer Vision related tasks like object detection, classification, unsupervised learning, and 3D scene understanding. At MEGVII, I'm mentored by Zeming Li and Dr. Gang Yu, under the supervision of Dr. Jian Sun.

I got my B.Eng. from Software Engineering College (Excellent Engineer Program) of South China University of Technology (one of the “Project 985” of China) with the First-Class Honor in July 2018. And I will become a PhD student at the Chinese University of Hong Kong (CUHK) in Fall 2021.

In my leisure time, I like to travel and take photos of the beautiful world. And I have a pet cat called PiPi.

Email  /  Google Scholar  /  CV  /  Github  /  LinkedIn  /  ZhiHu  /  Unsplash

profile photo
News
  • 2020.10 "EqCo: Equivalent Rules for Self-supervised Contrastive Learning" is available on arXiv.
  • 2020.07 The multi task computer vision toolkit cvpods is released! It supports classification, unsupervised learning, detection, segmentation and keypoints. Stay tuned for updates!!!
  • 2020.07 "AutoAssign: Differentiable Label Assignment for Dense Object Detection" is available on arXiv.
  • 2019.08 World's first general 3D object detection codebase Det3D is released!
  • 2019.06 - We win the nuScenes 3D object Detection challenge of WAD, CVPR 2019, by a large margin.
  • 2019.02 I join MEGVII Research.
  • 2019.01 My work - the LiDAR sensing demo is demonstrated on CES 2019.
  • 2018.06 I graduate from SCUT with the first class honor.
  • 2018.04 I join Horizon Robotics as a Perception Algorithm Engineer.
  • 2017.11 I join Alibabai as an intern Machine Learning Engineer.
  • 2017.07 I join DiDi AI Labi as an intern speech recognition engineer.
Selected Publications

My research main lies in Computer Vision and Robotics: 2D/3D Object Detection, Classification, Self-supervised Learning, 3D Scene Understanding. Representative papers are listed below. More info of publications can be found at Google Scholar.

EqCo: Equivalent Rules for Self-supervised Contrastive Learning
Benjin Zhu*, Junqiang Huang*, Zeming Li, Xiangyu Zhang, Jian Sun,
arXiv, 2020
code

EqCo breaks the widely accepted cognition of large negative samples in self-supervised contrastive learning and propose equivalent rules for the InfoNCE loss.

AutoAssign: Differentiable Label Assignment for Dense Object Detection
Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, Jian Sun,
arXiv, 2020
code

A fully differentiable label assignment strategy, which requires little human knowledge and is appearance-aware. It achieves SOTA(52.1% AP) on MS COCO.

Class-balanced Grouping and Sampling for Point Cloud 3D Object Detection
Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, Gang Yu,
arXiv, 2019
code

Technical report of the 1st place solution to the nuScenes 3D Object Detection.

Codebase

Apart from research, I have built and maintained many codebases for the research in our group. Make research more systematic is also exciting.

Det3D

World's first general 3D Object Detection framework. Winner of the nuScenes 3D Object Detection challenge, WAD, CVPR 2019.

cvpods

A versatile and efficient codebase for classification, segmentation, detection, self-supervised learning, keypoints and 3D. Widely used at MEGVII Research.

Competitions & Awards
  • Winner of the nuScenes 3D Object Detection challenge, WAD, CVPR 2019
  • 3rd Place of the Lyft 3D Object Detection challenge, NIPS 2019
  • SCUT Honored Graduate, 2018
  • SCUT scholarship 2015, 2016, 2017
Activities
  • 2020.08 Talk "AutoAssign: Differentiable Lable Assignment for Dense Object Detection"at cvmart. [Live Playback]
  • 2019.06 Talk "Winner of the nuScenes 3D Object Detection Challenge" at the WAD, CVPR 2019
Services
  • Conference Reviewer: CVPR'21
Blogs