Research Projects

Light Field Super-Resolution Rendering from Camera Array 

Goal: Compute refocused super-resolution (SR) images from light field captured by a camera array. This work was submitted to IEEE ICASSP 2014, entitled "Analysis of the Effect of Calibration Error on Light Field Super-resolution Rendering." [pdf]

My contribution:

  • Built the calibration programs and environment for light field acquisition from camera array.
  • Proposed an SR rendering algorithm and introduced a deconvolution step for successful SR rendering.
  • Analyzed the effect of calibration error on the algorithm
  • Proposed another single-shot SR focal stack rendering algorithm combining an optimization SR computing approach with a one-time calibration process.
 
https://sites.google.com/site/chenyuhsu1017/research/refocus.png?attredirects=0
Single-shot SR refocusing

Augmented Reality T-shirt Design and Try-On System (TDTOS)

Goal: Design an FPGA-based real-time system for users to design their T-shirts in a virtual fitting room. This work won the International Championship at the 2012 Altera Innovate Asia Workshop and Design Contest. A paper was published in The 1st Asia-Pacific Workshop on FPGA Applications, entitled "TDTOS - T-shirt Design and Try-On System." [pdf] [news cover] [demo]

My contribution:

  • Developed image processing and noise-tolerant algorithms to achieve realistic try-on simulation.
  • Developed pattern recognition algorithms and efficient memory control for gesture-based UI control, user-designed logo deformation, and real-time computation.
https://www.youtube.com/watch?v=AruIv2AG7_g
Snapshot of the YouTube demo (I am the model on the right.)

Network (Media Access Control) Protocol Design

Multichannel Network Protocol Design

Goal: Design an multichannel protocol for large-scale Machine-to-Machine network. This work was accepted by IEEE WCMC 2013, entitled "An Adaptive Multichannel Protocol for Large-scale Machine to Machine (M2M) Networks." [pdf]

My contribution:

  • Designed a distributed algorithm to estimate the number of competing devices in real-time.
  • Derived the optimal negotiation duration and access probability of devices using a common control channel.
  • Proposed an adaptive multichannel protocol to address the scalability issue in M2M communications.

After the presentation in IWCMC, my work was selected for publication in the WCMC journal. In this journal paper [pdf], I further introduced asynchronous resource allocation to the protocol for more realistic traffic pattern and achieved 93% channel utilization.

Hidden Node Effect

Goal: Alleviate the decrease of utilization due to hidden node effect. This work has been finalizing for submission. [pdf]

My contribution:

  • Analyzed factors (transmission power, negotiation target, network topology, etc.) affecting the hidden node effect.
  • Designed the length of the back-off slots to reduce hidden node collision, achieving as high as 50% utilization increase compared to p-CSMA protocol.
  • Proposed a symmetric topology for analysis, derived its mathematic model, and developed a C++ simulation program.
https://sites.google.com/site/chenyuhsu1017/research/multichannel.png?attredirects=0
Multichannel network with a common control channel 

https://sites.google.com/site/chenyuhsu1017/research/Async_Resource_Allocation_v2.png?attredirects=0
Asynchronous resource allocation

https://sites.google.com/site/chenyuhsu1017/research/hidden_collision.png?attredirects=0
Hidden node collisions 

Course Projects

Data-driven v.s. Model-driven in Large Image Dataset

 Courses: Multimedia Analysis and Indexing & Machine Learning (both graduate-level courses)    

Goal: Compare image retrieval (data-driven) and machine learning (model-driven) techniques in object recognition on ImageNet database (1.26 million images) [slides].

  • Collected quantized SIFT feature codewords (Bag of Words representations) into histogram and applied PCA to reduce dimension.
  • Built an object recognition system using image retrieval techniques and experimented on the number of nearest neighbors for voting
  • Trained pairwise linear SVM classification models for 112 synonym sets.
  • Compared the recognition precision of the two approaches on different synonym sets.
                          https://sites.google.com/site/chenyuhsu1017/research/data-driven.png?attredirects=0 https://sites.google.com/site/chenyuhsu1017/research/model-driven.png?attredirects=0
    Left: Data-driven approach. Right: Model-driven approach.

Recognizing Panoramas

 Course: Digital Visual Effects (graduate-level course)    

Goal: Given a set of images, find all possible panoramas and stitch them together.

  • Implemented feature detection, feature matching, image matching, bundle adjustment and blending.
  • Successfully recognized two sets of panoramas and stitched them together.
    https://sites.google.com/site/chenyuhsu1017/research/EEII_panorama.png
    The artifact, photographed at the roof top of EE-II building, National Taiwan University

Injury Prediction of Professional Basketball Players

 Course: Data Mining (graduate-level course)    

Goal: Given the players' statistic data, predict the player's tendency of injury in a period of time.

  • Defined Potential Injury Window (PIW) and mined the relation between the statistic data in this period and the injury time, considering the relation of sequential events as well.
  • Parsed statistic data from sport websites and trained logistic regression models using Weka.
  • Studied the way to avoid over-fitting by reducing the feature dimension.

Subscriber - A Smart Website Information Update Manager

Goal: Developed a Chrome application for users to manage updated information on any subscribed websites

  • Designed a user-friendly drag-and-drop interface for users to select DOM elements on webpages using JavaSciprt.
  • Developed a server that sends crawlers to subscribed websites periodically and parses the data to find information update.
  • Arranged users' updates on their personalized homepages.
 https://sites.google.com/site/chenyuhsu1017/research/subscriber.png
This project won the Best Popularity Award in the NTU innovation and Creation Contest, the largest venture competition held by NTU.