Hi there,

I'm Jing Bi

I love

design machine learning computer vision autonomous machine overwatch picking berries

About

A bit words about me

If I can summarize who I am in one word, that would be curious . I'm the kind of person that feels most comfortable when things are certain but feels most alive when they're not. That satisfaction of being able to learn something new(like building this website from scratch) is always a big thing that gets me up in the morning.

As for where I am, I'm now a research staff at University of Rochester, working on computer vision and autonomous control. My current work involves predicting and recognizing human intention and action in the egocentric video.

I've been working with Prof. Chenliang Xu. on the research project that supports autonomous system mimic behavior from human experts' demonstration, which facilitates teaching machine complex tasks with minimal prior knowledge. Before joining UofR, I was an Undergraduate Research Assitance at Shandong University, doing research on autonomous UAV and real-time tracking tasks with Prof.Wei Zhang

Projects

Research and some selected projects

Safe Imitation learning via self-prediction

Video Prediction

Learn from intervention

Text detection and reconigtion

incremental learning via VAE

General Object Tracking with UAV

Auto control UAV

3D Laser Constuction

Learning from interventions with hierarchical policies for safe learning

We propose a novel algorithm which combines Learning from Interventions with Hierarchical Imitation Learning. We also propose an interpolation trick called, Backtracking, that allows us to use state-action pairs before and after the intervention. Besides, a Triplet-Network based architecture which is capable of training the hierarchical policies.

2019
July - Sept.

Video Prediction

In this period, We implemented and evluated comparative evaluation among state-of-the-art Sequence Prediction and Generativemodels including ConvLSTM-2D/3D, Pred-Net/RNN++, VAE-GAN variations, and Pixel-RNN etc.

Besides, We bring the idea of deformable convolution into this project, and developed a motion matrix that contains normal image transformation function with learn parameter for disentanglemotion and content in moving video sequence. At last, we designed autoregressive 3d-Wavenet model and achieved competitive result with higher parallel performance.

2019
Jan. - June

Navigation by Imitation in a Pedestrian-Rich Environment

In this work, we introduced an Interactive learn-from-intervention algorithm with hierarchical framework for safer and faster learningthe control policy from real-world human behavior.

Besides, we built physical structure of an outdoor autonomous car with Nvidia TX2 and Arduino, and designed control and recording software. The proposed approach isApplied to a robotic vehicle in real-world campus sidewalk settingand the result indicates our method is data efficient approach for generalizing agood decision while ensuring safety

2018
May - Sept.

General Text Detection and Recognition

Implemented connectionist text proposal network for general text detection instead of using different size ofbounding box to scan whole image for multiple times+Designed and developed a end-to-end Conv-LSTM network for text recognition, which evenly slices detected imagepatch and extract the feature as input to sequence network.+Compared with classic general text recognition algorithm, The final pipline consists of both detection and recognition network and achieves superior competitive performance while maintaining fast inference tim

2018
March - May

Incremental learning with VAE

we analyzed the current state of the art method for class incremental learning, iCaRL, in detail, and showed that the primary reason iCaRL works well is not because of the herding algorithm. We also proposed a simple yet effective threshold moving method to fix the bias introduced by the distillation loss, and verified the effectiveness of this method. Finally, we use VAE to dis- cover the underlying feature distribution and use it to select representative samples, which help the distillation part obtain more meaningful samples form the previous trained dataset.

2018
Sept. - Dec.

Autonomous General Object Tracking with Unmanned Aerial Vehicles (UAVs)

In this work, we first proposaled a new framework to track general object with UAV. By combining a regression model for estimating object velocity using camera and drone dynamic measurement and a Siamese network based structure that is used to taking estimated distance of moving object as conditional input,The final tracking system is able to track assigned general moving objects offline in real-world with 30FPS.

2015 - 17
Sept. - June.

Auto UAV control

We rebuilt control and communication system with C++ based on ROS platform which is flexible for further extension. The final system can take waypoints or desired drone dynamic as input and execute action autonomously. This project is mainly an Engineering work served for tracking projects.

2016
May. - Sept.

3D laser construction

This is essentially an experimental project that serves as a building block for SLAM, which can estimate UAV's location indoor without GPS.

2016
Sept. - Dec.

Others

Services and a little selected awards

Teaching Assistance

Review

Award