Qieyun Dai

I'm a forth year PhD student in the Department of Computer Science, University of Illinois at Urbana-Champaign. I'm advised by Derek Hoiem.

Before joining UIUC, I received a Bachelor of Engineering from Nanjing University, China.

My research interests are Computer Vision and Machine Learning.


Ph.D. Candidate, Department of Computer Science, University of Illinois at Urbana-Champaign   Aug. 2009 – Present
Advisor: Derek Hoiem
GPA: 4.0/4.0
Bachelor of Engineering, Software Institute, Nanjing University   Sept. 2005 – June 2009
GPA: 92/100
Ranking: 1/213


Diagnosing Error in Object Detectors Diagnosing Error in Object Detectors
D.Hoiem, Y. Chodpathumwan and Q. Dai
ECCV 2012. (Oral Paper)
Paired Regions for Shadow Detection and Removal
R. Guo, Q. Dai, and D. Hoiem
TPAMI 02 Oct 2012.
Learning to Localize Detected Objects
Q. Dai and D. Hoiem
CVPR 2012.
Single-Image Shadow Detection and Removal using Paired Regions
R. Guo and Q. Dai and D. Hoiem
CVPR 2011. (Oral Paper)


• Object Localization

This project proposes an approach to accurately localize detected objects. The goal is to predict which features pertain to the object and define the object extent with segmentation or bounding box.

We use as our initial object detector a slight modification of the DPM detector by Felzenszwalb et al. that fires mostly within the object but is not required to cover the whole object.

We determine whether each pixel belongs to the object or not by building a color model that takes into account both class-level and instance level information, as well as classifying each edge pixel as either inside/outside/on the boundary of an object

Finally, we propose two different approaches to accurately localize the object: learned graph cut segmentation that accounts for both local cues and smoothness constraints and structural bounding box prediction that makes use of aggregated information.

• Shadow Detection and Removal

This project addresses the problem of shadow detection and removal from single images of natural scenes. Different from traditional methods that explore pixel or edge information, we employ a region based approach.

In addition to considering individual regions separately, we predict relative illumination conditions between segmented regions from their appearances and perform pairwise classification based on such information. We determine whether a pair of regions have the same material and illumination and use classification results to build a pairwise markov random field.

The labeling of shadow/non-shadow regions is solved using graph-cuts and refined by image matting. We define a lighting model where each pixel is lit by both the direct light and environment light and generate the shadow free image by relighting each pixel based on our lighting model.


ACM International Collegiate Programming Contest: Mid-Central USA Regional Champion (Representing U. Illinois), 2009
ACM International Collegiate Programming Contest: World Final Honorable Mention (Representing U. Illinois), 2010

Industrial Experience

Facebook Inc, Palo Alto, CA. Machine learning for spammy user requests detection, Summer 2010.
Google Inc, Mountain View, CA. Machine learning for click fraud detection, Summer 2011.
Google Inc, Mountain View, CA. Feature Developing for video classification, Summer 2012.

Professional Activities

Reviewer for Transaction of Graphics, ACCV 2012 and CVPR 2013

Graduate Courses

• Machine Learning
• Computer Vision
• Optimization in Machine Learning and Computer Vision
• Probabilistic Graphical Models