Yifei Liu

Hello World, I am a Research Assistant in ETH AIT Lab led by Prof. Otmar Hilliges with a current research focus on 3D human-object interaction reconstruction from a single in the wild image.

Before that, I did my master thesis at Robotics and Perception Group @ University of Zurich led by Prof. Davide Scaramuzza and worked on Computer Vision and Deep Learning with a focus on efficient inference and sparse computation, by means of pruning redundant tokens for Vision Transformers.

Even before, I obtained my BSc degree in Mathematics @ University of Science and Technology of China.

news

Oct 26, 2023	Our paper Revisiting Token Pruning for Object Detection and Instance Segmentation is accepted by IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2024.
May 1, 2023	I started a new position as a Research Assitant in AIT lab @ ETH, and will work on 3D human-object interaction reconstruction from in the wild images.
Sep 16, 2022	I started my masters thesis in Robotics and Perception Group focusing on Sparse Computation with Vision Transformers.
Jun 20, 2022	Our semester’s course project An Offline Python SLAM using COLMAP in the 3D Vision Course at Computer Vision and Geometry group in ETH got full-grade and presented in the poster session!
Sep 7, 2021	I joined the Robotics and Perception Group as a master student on project Efficient Spatio-Temporal Processing of Event Data.
Jul 3, 2021	We got the 2nd place in the Competition of 3D Object Detection based on Lidar Points from the ETH Course Deep Learning for Autonomous Driving
Jun 15, 2021	I win the 1st place in the final written coding exam in Informatik II among more than 250 students
Sep 1, 2020	I started my masters study in University of Zurich (UZH) with a major of Data Science and minor of Informatics.

Publications

Revisiting Token Pruning for Object Detection and Instance Segmentation
Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, Davide Scaramuzza

[PDF] [Code] [Cite] [Project]

Pruning tokens for Vision Transformers to accelerate inference speed on dense tasks.

Projects

	An Offline Python SLAM using COLMAP Semester's Course Project in 3D Vision , Computer Vision and Geometry Group, ETH Zurich Supervised by Paul-Edouard Sarlin and Marc Pollefeys Webpage \| Poster \| Demo \| Report TL;DR: An offline python SLAM using COLMAP and ORB-SLAM which is robust, accurate, and highly extensible.
	Efficient Spatio-Temporal Processing of Event Data Master Project at Robotics and Perception Group, UZH Supervised by Mathias Gehrig, Nico Messikommer and Davide Scaramuzza Webpage \| Slide \| Report TL;DR: A point-voxel-mixed model for efficient spatio-temporal processing of event data for classification and optical flow.
	Multi-task Learning for Autonomous Driving Course Project in Deep Learning for Autonomous Driving Webpage \| Competition Ranking \| Report Multi-task Learning for semantic segmentation, depth estimation and 3D object detection for autonomous driving scenes.
	Collection of Demos Here are some demos collected from my courses. Please note that I am not an expert in these fields, and am merely recording some experience I found interesting to share.

	An Offline Python SLAM using COLMAP Semester's Course Project in 3D Vision , Computer Vision and Geometry Group, ETH Zurich Supervised by Paul-Edouard Sarlin and Marc Pollefeys Webpage \| Poster \| Demo \| Report TL;DR: An offline python SLAM using COLMAP and ORB-SLAM which is robust, accurate, and highly extensible.
	Efficient Spatio-Temporal Processing of Event Data Master Project at Robotics and Perception Group, UZH Supervised by Mathias Gehrig, Nico Messikommer and Davide Scaramuzza Webpage \| Slide \| Report TL;DR: A point-voxel-mixed model for efficient spatio-temporal processing of event data for classification and optical flow.
	Multi-task Learning for Autonomous Driving Course Project in Deep Learning for Autonomous Driving Webpage \| Competition Ranking \| Report Multi-task Learning for semantic segmentation, depth estimation and 3D object detection for autonomous driving scenes.
	Collection of Demos Here are some demos collected from my courses. Please note that I am not an expert in these fields, and am merely recording some experience I found interesting to share.