About Me

avatar

Before coming to UCSD, I spent four wonderful years at University of Science and Technology of China(USTC) to get a B.E in Electronic Engineering and Information Science(1006).

Education

ucsd_logo

University of California, San Diego

Sep 2014 - Nov 2017

Master of Science, Electrical and Computer Engineering

Overall GPA:3.945/4.0

ustc_logo
University of Science and Technology of China (USTC)
Sep 2010 - Jul 2014

BE, Electronic Engineering and Information Science

Overall GPA:4.02/4.3 Rank:2/133

Projects

gesture
Large-scale Isolated Gesture Recognition Challenge
Aug 2017 - Sep 2017, Chalearn LAP Challenge@ICCV 2017
  • Challenged to recognize 249 classes of gestures from RGB + D (depth) videos
  • Trained a multi-modal 3D-CNN with more than 50,000 samples for gesture recognition
  • Proposed 1) Region of Interest Masking 2) Multi-modality finetuning 3) Spatial-Temporal Pyramid encoding to improve performance
  • Ranked 3rd place out of 12 attending teams, code and model is available.
tag
Image Tagging and Search with Image-text Embedding
June 2016 - Sep 2016, Adobe Research, Research Intern
  • Improved the in-house image auto-tagging and search system with deep learning.
  • Trained an image embedding network with PMI word vectors
  • Boosted the image search system with click-through data by ranking with positive enhancement
  • Derived an online update algorithm for PMI word embedding with a large dictionary
  • Applied the embedding network for image dense tagging in an online web demo
flapper
Playing FlappyBird with Reinforcement Learning
Apr 2017 - Jun 2017, Side Project
  • An interesting trial to apply reinforcement learning on playing games: video link
  • Implemented the game simulator to interact with the reinforcement learning model
  • Designed a Deep Q-learning model with memory replay
  • Compared with a hand-crafted baseline policy and the Deep Q-learning achieved better results video link
vlad3
VLAD3: Encoding Dynamics of Deep Features for Action Recognition
Jul 2015 - Nov 2015, Research Project
  • Studied the importance of dynamics modeling in video action recognition
  • Derived a VLAD encoding(VLAD3) with Linear Dynamic System(LDS) model for deep feature sequence
  • Implemented the codebook learning and encoding of VLAD3 with a modified Kalman smoothing algorithm
  • Benchmarked against common baselines and achieved superior performance on common benchmark datasets(e.g. Olympic Sports, UCF101 and THUMOS14)
map3d
Building Large Scale 3D Map with Kinect
Dec 2013 - May 2014, Microsoft Research Asia, Research Intern
  • Designed and implemented an indoor 3D map reconstruction system: video link
  • Input: RGBD data from Microsoft Kinect sensor
  • Process I: Registration of consecutive frames by visual feature matching
  • Process II: Globally optimize the pose graph and align large planes
  • Process III: De-noise and re-assign color of 3D points.
  • Output: Full-size 3D map of a large indoor scene.