Welcome!

Hi, I'm KUNG-HUNG LU.

I'm a master student studying Electrical and Computer Engineering at University of Washington.
My interests lies in solving real-world problems related to Robotics and Artificial Intelligence.
In particular, I am interested in understanding the way people perceive and interact with various data, and finding ways to harness the tools of machine learning to enable machines to solve complicated tasks.

Skills

Programming Languages

C/C++, Python, Java, Matlab, Html, Javascript

Tools

Qt, OpenCV, Point Cloud Library (PCL), ROS,
scikit-learn, Caffe, pyTorch, Tensorflow

Robotics

  1. Localization and Mapping
  2. Reinforcement Learning

Computer Vision

  1. 3D perception
  2. Image Captioning
  3. Semantic Segmentation

Projects

Aesthetic Critiques Generation for Photos

Kung-Hung Lu*, Kuang-Yu Chang*, and Chu-Song Chen, “Aesthetic Critiques Generation for Photos,” International Conference on Computer Vision (ICCV), October 2017 (* indicates equal contributions).
We developed a CNN-LSTM model to generate abundant aesthetics captions, compared to common image captioning tasks that depict the objects or their relations in a picture. The results of experiments on PCCD demonstrate that our approaches outperform existing methods for generating aesthetic-oriented captions for images.

Monocular Visual Odometry

Estimating the motion of a camera in real time is a critical module for many robotics applications such as self-driving car. In this project, I implemented a monocular visual odometry from scratch with only C++, openCV and G2o. It utilized ORB feature to do scene matching and perspective-N-points algorithm to do pose estimation. Our method also maintained a local map to improve the performance and accelerate the process.

Comparison of Particle Based Filtering Algorithms in Localization

We studied localization approaches and provide an in-depth, comprehensive evaluation of different localization algorithm. Specifically, we implemented different variations of the Particle Filter, with a focus on investigating various resampling methods. Then we evaluated each algorithm using ROS simulator and compare the performance with standard metrics used in the literatures.

Deep Q Learning on Atari Game - Breakout

In this project, we show the deep reinforcement learning has a good ability to learn how to play the game Breakout. We uses a convolutional neural network to approximate a reward function, Q-function to predict the longterm reward to guide the decision making process and demonstrate that it is able to achieve human-level performance. Furthermore, We also exploit the target network and double Q network to alleviate the slow convergence problem. The experiment results show that the training process indeed speed up with these improvements.