News register
Search section
Search area
News Type
News type
Search date
Search word OR
List of related articles
Contents of related articles
No info was found
List of related articles
Contents of related articles
No info was found
View details

04/30/2019 HYU News > Academics > 이달의연구자


[Researcher of the Month] Hedging Deep Features for Visual Tracking

Professor Lim Jong-woo (Division of Computer Science and Engineering)


Copy URL / Share SNS

Professor Lim Jong-woo (Division of Computer Science and Engineering) has recently published his thesis "Hedging Deep Features for Visual Tracking" in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI). The thesis was mainly about discovering an algorithm that enables visual tracking of objects' features within a video through an artificial neural network.
Professor Lim Jong-woo (Division of Computer Science and Engineering) wrote a thesis on discovering an algorithm that enables the visual tracking of the features of objects within a video through an artificial neural network.

Convolutional Neural Network (CNN)

According to Lim, the studies of Convolutional Neural Networks (CNN) have led to the development of the field of computer vision and deep learning by producing continuous results over the past 10 years. A CNN refers to a network that delivers information about certain objects through the analysis of data produced within each layer of information. The former part of the network provides information about the location of a certain object, whereas the latter mainly handles data related to the type or meaning of the object. In order to accurately position or track an object within a video, both types of information are required, which leads to the need for technology that can successfully fuse the provided information within each layer.

Lim’s recent research enables this particular fusion of the various layers of information by applying a technology called hedging, which was derived from machine learning methods. Whenever a new frame is input, the location of the object is traced based upon a corrective filter that collects the information within the different layers of the CNN. This "history sensitive hedging" method, which was newly introduced in Lim’s studies, recalls every result of each layer, enabling the selection of the most relevant layer in relation to the present frame. This hedging method has allowed more efficient and productive results compared to the conventional methods of the visual tracking of objects.
The overall visual tracking process that occurs within a newly input frame, using the hedging technology introduced by Lim
(Photo courtesy of Lim)

The significance of the research

Lim’s recent thesis is an extension of his past research "Hedged Deep Thinking," published in 2016 in the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), which focused on applying the visual features in deep learning to the procedures for tracking objects. The conventional method was simply using all the positional information derived from each layer of the CNN, yet the CVPR thesis introduced an algorithm that selects only the necessary information through a hedging method.

“My latest thesis is one that expands such studies by applying a Siamese network that further distinguishes the object of scrutiny and a cumulative regret model that reflects the former weight measures within previous frames into the most present one,” explained Lim. A Scale search step was also added in order to prepare for situations in which the size of the object changes. Lim stated that his newly presented algorithm has allowed a further step within the field of visual tracking technology by producing superior results compared to the conventional State-of-the-art algorithm.

For an easier understanding, Lim pointed out the current CCTV business. Enabling the visual tracking of certain objects within a video, which includes both organic and inorganic materials, Lim’s newly introduced algorithm allows this process to be conducted in a more intelligent manner by efficiently selecting only the necessary features within each layer of the CNN. As an example, Lim illustrated a potential use of this new technology by using the CCTV monitoring process at airports. His new technology would make the process of identifying and tracking suspicious figures easier. Automatic driving and action cameras were also other fields that Lim presented as examples the technology could be applied to.
Lim is explaining the significance of his newly published thesis and how it can be applied to more practical fields.

Future plans

Lim shared his future plans of conducting research on the multiplex visual tracking technology that allows the tracking process to be widened towards multiple objects, which is an expansion of the current visual tracking technology which mainly focuses on tracking a single object. Furthermore, he also stated how he and his laboratory students are working on the image-based attitude estimation and three dimensional restoration technologies, which can be related to the fields of automatic driving, AR/VR, and even robots.

As for the last comments, Lim stated that many professors and students of the Division of Computer Science and Engineering are currently conducting research on artificial intelligence and its related fields such as computer vision and data mining. “I hope that such research aids in the needed progress in this field, and that the professors and students are able to find various opportunities of acquiring the professional knowledge to develop their strengths and abilities,” finished Lim.   

Choi Seo-yong
Photos by Kim Joo-eun
Copy URL / Share SNS