Wei Tong

Wei Tong

Postdoctoral Researcher

Language Technologies Institute
Carnegie Mellon University

Email: tongwei AT cs DOT cmu DOT edu

I received my Ph.D. degree from Michigan State University in 2010 and my advisor was Dr. Rong Jin. I am currently working as a postdoc in the Language Technologies Institute , Carnegie Mellon University .

Research Interests:

Large scale content-based image/video retrieval
Machine learning
Convex optimization
Recommender systems
Autonomous vehicle navigation

Research Projects:

	Automatic Graffiti Matching and Gang/Moniker Identification Graffiti are any type of public markings that may appear in forms ranging from simple written words to elaborate wall paintings. It has existed since ancient times and now graffiti are abundant in most urban neighborhoods. Identifying criminal gangs and monikers is one of the most important tasks for graffiti analysis in law enforcement. In current practice, this is typically performed manually by the law enforcement officers, which is not only time-consuming but also results in limited identification performance. In this project we design and develope an system which can automatically match the given graffiti against the graffiti database and identify the gang or moniker of the given graffiti. The challanges for graffiti matching and gang/moniker identification are three folds. First, there are a limited number of graffiti images that are associated with each gang/moniker, making it difficult to directly apply the standard supervised learning techniques. Second, due to the large variance in the visual appearance of the graffiti generated by the same gang and moniker, it is not easy to find the matched graffiti images using the standard image retrieval algorithms. Finally, due to the freestyle of graffiti, the textual content of graffiti cannot always be recognized, making it difficult to explore text categorization and retrieval algorithms.

	Automatic Tattoo Image Matching and Retrieval Tattoo images on the human body have been routinely collected and used in law enforcement to assist in suspect and victim identification. However, the current practice of matching tattoos is based on keywords. Assigning keywords to individual tattoo images is both tedious and subjective. In this project, we are developing a content-based image retrieval system for a large scale tattoo image database. The system utilizes local image feautes to automatically retrieve database imagese which have similar visual contents to the given query tattoo image. Side information, i.e., body location of tattoos and tattoo classes, is utilized to improve the retrieval time and retrieval accuracy. Geometrical constraints are also introduced in keypoint matching to reduce false retrievals.

	A Kernel Density Based Approach for Large Scale Image Retrieval In recent studies of Content-Based Image Retrieval (CBIR), most existing approaches represent an image by a bag-of-words model in which every local feature is quantized into a visual word. Given the bag-of-words representation for images, a text search engine is then used to efficiently find the matched images for a given query. The main drawback with these approaches is that the two key steps, i.e., key point quantization and image matching, are separated, leading to sub-optimal performance in image retrieval. In this work, we present a statistical framework for large-scale image retrieval that unifies key point quantization and image matching by introducing kernel density function. The key ideas of the proposed framework are (a) each image is represented by a kernel density function from which the observed key points are sampled, and (b) the similarity of a gallery image to a query image is estimated as the likelihood of generating the key points in the query image by the kernel density function of the gallery image. We present efficient algorithms for kernel density estimation as well as for effective image matching. Experiments with large-scale image retrieval confirm that the proposed method is not only more effective but also more efficient than the state-of-the-art approaches in identifying visually similar images for given queries from large image databases.

	An Efficient Key Point Quantization Algorithm for Large Scale Image Retrieval Recent studies have shown that local image features, are effective for near duplicate image retrieval. The most popular approach for key point based image matching is the clustering-based bag-of-words model. It maps each key point to a visual word in a code-book that is constructed by a clustering algorithm, and represents each image by a histogram of visual words. Despite its success, there are two main shortcomings of the clustering-based bag-of-words model: (i) it is computationally expensive to cluster millions of key points into thousands of visual words; (ii) there is no theoretical analysis on the performance of the bag-of-words model. We propose a new scheme for key point quantization that addresses these shortcomings. Instead of clustering, the proposed scheme quantizes each key point into a binary vector using a collection of randomly generated hyper-spheres, and a bag-of-words model is constructed based on such randomized quantization. Our theoretical analysis shows that the resulting image similarity provides an upper bound for the similarity based on the optimal partial matching between two sets of key points. Empirical study on a database of 1,000, 000 images shows that the proposed scheme is not only more efficient but also more effective than the clustering-based approach for near duplicate image retrieval.

	Second-Order PCA-Style Dimensionality Reduction Algorithm by Semidefinite Programming Principal Component Analysis (PCA) is a widely used dimensionality reduction method in computer vision, pattern recognition and machine learning. It is however computationally expensive when the dimension of data is very high. Recently, several Second-order PCA-style (SOPCA) dimensionality reduction algorithms have been proposed in order to overcome the computational difficulty in applying PCA to images which are conventionally viewed as vectors of high dimensionality. The main idea behind these algorithms is to avoid converting an image into a high dimensional vector but use its natural matrix representation directly. However, most SOPCA algorithms require solving a non-convex optimization problem, and therefore are only able to identify local optimal solutions. In this work, we propose a convex formulation for SOPCA that can eventually be represented as a Semi-defnite Programming (SDP) problem which is convex. An approximate algorithm is also presented to efficiently solve the related SDP problem. A theoretical analysis shows a very close relationship between proposed SDP fomulation and the common non-convex formulation.

	Semi-supervised Learning by Mixed Label Propagation The key idea behind many graph-based semi-supervised learning approaches is to enforce the consistency between the class assignment of unlabeled examples and the pairwise similarity between examples. One major limitation with most graph-based approaches is that they are unable to explore dissimilarity or negative similarity. This is because the dissimilar relation is not transitive, and therefore is difficult to be propagated. Furthermore, negative similarity could result in unbounded energy functions, which makes most graphbased algorithms unapplicable. In this paper, we propose a new graph-based approach, termed as “mixed label propagation” which is able to effectively explore both similarity and dissimilarity simultaneously. In particular, the new framework determines the assignment of class labels by (1) minimizing the energy function associated with positive similarity, and (2) maximizing the energy function associated with negative similarity.

	Autonomous Vehicle Navigation Team member of Autonomous Vehicle System for DARPA Grand Challenge 2005. I worked with other team members on obstacle avoidance (using LIDARs and cameras), GPS positioning and path planning. The team was one of the 40 teams entered the NQE (National Qualification Event)

	Developmental Robots and Autonomous Mental Development Traditional approaches to machine intelligence require human designers to explicitly program task-specific representation, perception and behaviors, according to the tasks that the machine is supposed to execute. However, AI tasks require capabilities such as vision, speech, language, and motivation which have been proved to be too muddy to program effectively by hand. Autonomous development is the nature's approach to human intelligence. The goal of this line of research is to understand human intelligence and to advance artificial intelligence. For the former, we need to understand how the brain acquires its rich array of capabilities through autonomous development. For the latter, we aim at machine's human-level performance through autonomous development. Developmental programs, such as SHM, IHDR, MILN, etc., was applied for a robot to develop its mind through the learning process.

Publications:

Papers:

1.	Jun Yang, Wei Tong, Alexander G. Hauptmann, "A Framework for Classifier Adaptation for Large-Scale Multimedia Data", Proceedings of the IEEE, 100(9): 2639-2657, 2012
2.	Wei Tong, Fengjie Li, Rong Jin and Anil Jain, "Large-scale near-duplicate image retrieval by kernel density estimation", International Journal of Multimedia Information Retrieval, 1(1): 45-58, 2012
3.	Yang Cai, Wei Tong, Linjun Yang, Alexander Hauptmann, “Constrained Keypoint Quantization: Towards Better Bag-of-Words Model for Large-scale Multimedia Retrieval”, ACM International Conference on Multimedia Retrieval, 2012
4.	Wei Tong, Jun-Eun Lee, Rong Jin, and Anil K. Jain, "Gang and Moniker Identification by Graffiti Matching", International ACM Workshop on Multimedia in Forensics and Intelligence (MiFor), 2011
5.	Jun-Eun Lee, Wei Tong, Rong Jin, and Anil K. Jain "Image Retrieval in Forensics: Application to Tattoo Image Database", IEEE Multimedia, 2011
6.	Wei Tong, Fengjie Li, Tianbao Yang, Rong Jin, Anil Jain. A Kernel Density Based Approach for Large Scale Image Retrieval. In ACM International Conference on Multimedia Retrieval, 2011
7.	Wei Tong, Tiaobao Yang , Rong Jin. Co-training For Large Scale Image Classification: An Online Approach. In ICPR workshop on Analysis and Evaluation of Large-Scale Multimedia Collections, 2010
8.	Wei Tong, Rong Jin. Second-Order PCA-Style Dimensionality Reduction Algorithm by Semidefinite Programming. In Snowbird Learning Workshop, 2009
9.	Wei Tong, Rómer Rosales, Glenn Fung, Automatic discrimination of mislabeled training points for large margin classifiers. In Snowbird Learning Workshop, 2009
10.	Fengjie Li, Wei Tong, Rong Jin, Anil Jain. An Efficient Key Point Quantization Algorithm for Large Scale Image Retrieval, In ACM Multimedia International Conference Workshop on Large-scale Multimedia Retrieval and Mining, 2009
11.	Wei Tong, Rong Jin. Semi-supervised Learning by Mixed Label Propagation. In The Twenty-First AAAI Conference on Artificial Intelligence, 2007
12.	Zhengping Ji, Xiao Huang, Wei Tong and John Weng, “On-line Learning of Covert and Overt Perceptual Capability for Vision-based Navigation”, In IEEE International Conference on Development and Learning (ICDL’06), 2006
12.	Wei Tong, Huawen Sun, Daling Wang, etc. Study and Implementation on a Middleware Technique for Web Personalized Recommendation Service (in Chinese). In National Database Conference, China, 2003.
14.	Shuangshuang Xie, Wei Tong, Yubin Bao, etc. Study On Data Mining Process Management Based On Workflow (in Chinese). In National Database Conference, China, 2002.
15.	Daling Wang, Yubin Bao, Ge Yu, Guoren Wang, Personalized Learning Systems in Distance Education. In Pan-Yellow-Sea International Workshop on Information Technologies for Network Era, 2002

Textbooks:

Practical Data Structure (in Chinese), Wei Tong, Shuangshuang Xie, Science Press, China, 2003, ISBN: 7-03-010930-9/TP.1860

Industrial Experience:

05/2008 – 08/2008	Student summer intern in CAD and Knowledge Solutions (IKM CKS), Siemens Medical Solutions Inc, USA
05/2007 – 08/2007	Student summer intern in Research & Development Department, General Motors

Teaching Experience:

09/2005 – 12/2005	Teaching assistant for CSE 231 "Introduction to Programming"
09/2004 – 05/2005	Teaching assistant for CSE131 "Introduction to Technical Problem Solving with Computer"

Working Experience:

01/2006 – 12/2008	Computer system administrator of Institute of Water Research, MSU
07/2000 – 07/2001	Agriculture Bank of China

Last modified: 09/2012