Olsen Hall Room 305, University of Massachusetts Lowell
One University Avenue, Lowell, MA 01854 USA
Email: bsun AT cs DOT uml DOT edu
I am currently a PhD student at the Computer Vision and Machine Learning Group of University of Massachusetts Lowell and my advisor is Prof. Kate Saenko. I got my MS from University of Massachusetts Lowell and BEng from Shandong University, both in Computer Science.
NEW! Our paper titled Return of Frustratingly Easy Domain Adaptation (Extended Abstract) has won the Best Paper Prize of the Transferring and Adapting Source Knowledge in Computer Vision (TASK-CV) workshop at ICCV 2015.
NEW! I am co-organizing NIPS 2015 Workshop on Transfer and Multi-Task Learning: Trends and New Perspectives.
Computer Vision, Machine Learning, Domain Adaptation, Deep Learning
 B. Sun, J. Feng, and K. Saenko. Return of Frustratingly Easy Domain Adaptation. In AAAI, 2016 [Preprint]
 X. Peng, B. Sun, K. Ali, and K. Saenko. Learning Deep Object Detectors from 3D Models. In ICCV, 2015 [Paper]
 B. Sun, K. Saenko. Subspace Distribution Alignment for Unsupervised Domain Adaptation. In BMVC, 2015 [Paper]
 B. Sun, K. Saenko. From Virtual to Reality: Fast Adaptation of Virtual Object Detectors to Real Domains. British Machine Vision Conference (BMVC), 2014 [Project page]
 B. Sun, X. Sun, Y. Wu, Y. Yin, and G. Yang. A New Algorithm for Generating Unique-Solution Sudoku. ICNC, 2008.
 A. Chakrabarti, Y. Xiong, B. Sun, T. Darrell, D. Scharstein, T. Zickler, and K. Saenko. Modeling Radiometric Uncertainty for Vision with Tone-mapped Color Images. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2014 [Project page]
 B. Sun, J. Feng, and K. Saenko. Return of Frustratingly Easy Domain Adaptation. ICCV workshop on Transferring and Adapting Source Knowledge in Computer Vision (TASK-CV), 2015 [Poster] (Best Paper Prize)
 B. Sun, K. Saenko. From Virtual to Reality: Fast Adaptation of Virtual Object Detectors to Real Domains. TASK-CV: Transferring and Adapting Source Knowledge in Computer Vision at ECCV 2014 [Poster]
 Correlation Alignment for Unsupervised Domain Adaptation. Google. Mountain View, CA, 2016.
 Correlation Alignment for Unsupervised Domain Adaptation. Google Cambridge. MA, 2016.
 Correlation Alignment for Unsupervised Domain Adaptation. Amazon. MA, 2016.
 Correlation Alignment for Unsupervised Domain Adaptation. Apple Inc. California, 2015.
 Correlation Alignment for Unsupervised Domain Adaptation. Vecna Technologies. Cambridge, MA, 2015.
 From Virtual to Reality: Fast Adaptation of Virtual Object Detectors to Real Domains. Amazon Fall Graduate Research Symposium. Seattle, 2014. [Slides]
Honors and Awards
Best Paper Prize, Transferring and Adapting Source Knowledge in Computer Vision (TASK-CV) workshop at ICCV 2015
PM Raj Fund for Excellence in Computer Science 2014
Outstanding Student Scholarship (top 5-10%) 2005 - 2009
Scientific and Technological Innovation Scholarship (top 0.5%) 2008
1st Prize, National Post-Graduate Mathematical Contest in Modeling (top 4.1%) 2008
Dean's Scholarship (top 0.5%) 2007
5th Place, Robocup China Open 2007 (5/40) 2007
1st Prize, China Undergraduate Mathematical Contest in Modeling (top 2.1%) 2007
University of Massachusetts Lowell, Lowell, MA 01/2013 - Present
Graduate Student Researcher with Prof. Kate Saenko
Researching on computer vision and machine learning
ICSI & EECS, UC Berkeley, Berkeley, CA 05/2015 - 08/2015
Researching on domain adaptation, deep learning
University of Massachusetts Lowell, Lowell, MA 09/2010 - 12/2012
Graduate Student Researcher with Prof. Georges Grinstein
Researching on data visualization and data mining
Shandong University, Jinan, China 01/2008 - 2010/06
Undergraduate and Graduate Student Researcher with Prof. Xinshun Xu
Researching on machine learning and computational Intelligence
University of Massachusetts Lowell, Lowell, MA 09/2010 - 05/2013
Teaching Assistant for Machine Learning, Organization of Programming Languages, Topics in Bioinformatics, Visual Analytics, Operating Systems, and Computer Architecture.
Correlation Alignment for Unsupervised Domain Adaptation
Unlike human learning, machine learning often fails to handle changes between training (source) and test (target) input distributions. Such domain shifts, common in practical scenarios, severely damage the performance of conventional machine learning methods. Supervised domain adaptation methods have been proposed for the case when the target data have labels, including some that perform very well despite being "frustratingly easy" to implement. However, in practice, the target domain is often unlabeled, requiring unsupervised adaptation. We propose a simple, effective, and efficient method for unsupervised domain adaptation called CORrelation ALignment (CORAL). CORAL minimizes domain shift by aligning the second-order statistics of source and target distributions, without requiring any target labels. Even though it is extraordinarily simple--it can be implemented in four lines of Matlab code--CORAL performs remarkably well in extensive evaluations on standard benchmark datasets. [AAAI 2016 Paper (preprint)]
Building ImageNet in One Day
Datasets power computer vision research and drive breakthroughs. Larger and larger datasets are needed to better utilize the exponentially increasing computing power. However, datasets generation is both time consuming and expensive as human beings are required for image labelling. Human labelling cannot scale well. How can we generate larger image datasets easier and faster? In this paper, we provide a new approach for large scale datasets generation. We generate images from 3D object models directly. The large volume of freely available 3D CAD models and mature computer graphics techniques make generating large scale image datasets from 3D models very efficient. As little human effort involved in this process, it can scale very well. Rather than releasing a static dataset, we will also provide a software library for dataset generation so that the computer vision community can easily extend or modify the datasets accordingly. [Project Page]
Exploring Invariances In Deep Convolutional Neural Networks Using Synthetic Images
Deep convolutional neural networks learn extremely powerful image representations, yet most of that power is hidden in the millions of deep-layer parameters. What exactly do these parameters represent? Recent work has started to analyze CNN representations, finding that, e.g., they are invariant to some 2D transformations, but are confused by particular types of image noise. In this project, we delve deeper and ask: how invariant are CNNs to object-class variations caused by 3D shape, pose, and photorealism? These invariance properties are difficult to analyze using traditional data, so we propose an approach that renders synthetic data from freely available 3D CAD models. Using our approach we can easily generate an infinite amount of training images for almost any object. We explore the invariance of CNNs to various intra-class variations by simulating different rendering conditions, with surprising findings. Based on these results, we propose an optimal synthetic data generation strategy for training object detectors from CAD models. We show that our Virtual CNN approach significantly outperforms previous methods for learning object detectors from synthetic data on the benchmark PASCAL VOC2007 dataset. [ICCV 2015 Paper] [ICLR 2015 Workshop Paper]
Probabilistic Derendering of Camera Tone-Mapped Images
To produce images that are suitable for display, tone-mapping is widely used in digital cameras to map linear color measurements into narrow gamuts with limited dynamic range. This introduces non-linear distortion that must be undone, through a radiometric calibration process, before computer vision systems can analyze such photographs radiometrically. This project considers the inherent uncertainty of undoing the effects of tone-mapping. We observe that this uncertainty varies substantially across color space, making some pixels more reliable than others. We introduce a model for this uncertainty and a method for fitting it to a given camera or imaging pipeline. Once fit, the model provides for each pixel in a tone-mapped digital photograph a probability distribution over linear scene colors that could have induced it. We demonstrate how these distributions can be useful for visual inference by incorporating them into estimation algorithms for a representative set of vision tasks. [Project Page]
From Virtual to Reality: Fast Adaptation of Virtual Object Detectors to Real Domains
The most successful 2D object detection methods require a large number of images annotated with object bounding boxes to be collected for training. We present an alternative approach that trains on virtual data rendered from 3D models, avoiding the need for manual labeling. Growing demand for virtual reality applications is quickly bringing about an abundance of available 3D models for a large variety of object categories. While mainstream use of 3D models in vision has focused on predicting the 3D pose of objects, we investigate the use of such freely available 3D models for multicategory 2D object detection. To address the issue of dataset bias that arises from training on virtual data and testing on real images, we propose a simple and fast adaptation approach based on decorrelated features. We also compare two kinds of virtual data, one rendered with real-image textures and one without. Evaluation on a benchmark domain adaptation dataset demonstrates that our method performs comparably to existing methods trained on large-scale real image domains. [Project Page] [Poster] [Slides]
Deconstructing the Deformable Parts Model: Do More with Less
The deformable parts model (DPM) is a successful detection model that continues to achieve state-of-the-art results in object detection. While the model's success is attributed to the deformable parts, the proposed system has many other design choices such as the use of multi-resolution HOG features, the sampling of parts from high energy areas, the use of a mixture model, etc. Recently, there have been attempts at analyzing the contribution of these design choices more closely. We delve further into this analysis, performing a study of the effects of restricting the deformation model, the use of single-resolution filters versus multi-resolution filters, and the application of energy 'dropout' while sampling parts. Our results indicate that a better performance can be achieved with simpler cost-free deformation models. [Paper] [Poster]
WEAVE: Web-based Analysis and Visualization Environment
Weave is a new web-based visualization platform designed to enable visualization of any available data by anyone for any purpose. Weave is an application development platform supporting multiple levels of users -- novice to advanced -- as well as the ability to integrate, analyze and visualize data at 'nested' levels of geography, and to disseminate the results in a web page. [Project Page]