Dongjin Kim
Hanyang University · Computer Science
About the Lab
Professor Dongjin Kim's research lab specializes in computer vision, artificial intelligence, and smart materials, with a strong focus on relational understanding in visual scenes, event-driven simulation, and energy-harvesting technologies. The lab develops advanced AI models for dense relational captioning and multi-task learning to enhance image and behavior understanding, while also pioneering novel materials and devices such as electroluminescent polymers and piezoelectric energy harvesters. Their work bridges fundamental AI research with practical applications in human-computer interaction, autonomous systems, and sustainable energy solutions. The lab emphasizes data efficiency, structural priors (like part-of-speech tagging), and real-time simulation for complex physical systems.
Research Overview
Research Output Trend
Figures are computed from collected data and may differ slightly.
Selected Papers
15Our goal in this work is to train an image captioning model that generates more dense and informative captions. We introduce "relational captioning," a novel image captioning task which aims to generate multiple captions with respect to relational information between objects in an image. Relational captioning is a framework that is advantageous in both diversity and amount of information, leading to image understanding based on relationships. Part-of-speech (POS, i.e. subject-object-predicate ca
This paper presents an event-driven approach that efficiently detects collisions among multiple ballistic spheres moving in the 3D space. Adopting a hierarchical uniform space subdivision scheme, we are able to trace the trajectories of spheres and their time-varying spatial distribution. We identify three types of events to detect the sequence of all collisions during our simulation: collision, entering, and leaving. The first type of event is due to actual collisions, and the other two types o
We introduce dense relational captioning, a novel image captioning task which aims to generate multiple captions with respect to relational information between objects in a visual scene. Relational captioning provides explicit descriptions for each relationship between object combinations. This framework is advantageous in both diversity and amount of information, leading to a comprehensive image understanding based on relationships, e.g., relational proposal generation. For relational understan
Human behavior understanding is arguably one of the most important mid-level components in artificial intelligence. In order to efficiently make use of data, multi-task learning has been studied in diverse computer vision tasks including human behavior understanding. However, multitask learning relies on task specific datasets and constructing such datasets can be cumbersome. It requires huge amounts of data, labeling efforts, statistical consideration etc. In this paper, we leverage existing si
ADVERTISEMENT RETURN TO ISSUEPREVCommunication to the...Communication to the EditorNEXTSynthesis of a New Class of Processable Electroluminescent Poly(cyanoterephthalylidene) Derivative with a Tertiary Amine LinkageDong-Jin Kim, Sung-Hyun Kim, Taehyoung Zyung, Jang-Joo Kim, Iwhan Cho, and Sam Kwon ChoiView Author Information Department of Advanced Materials Engineering, Korea Advanced Institute of Science and Technology, P.O. Box 201, Cheongryang, Seoul, Korea, Department of Chemistry, Korea Adv
We designed and fabricated a high performance spring-type piezoelectric energy harvester that selectively collects current from the inner part of a spring shell.
A common problem in the task of human-object interaction (HOI) detection is that numerous HOI classes have only a small number of labeled examples, resulting in training sets with a long-tailed distribution. The lack of positive labels can lead to low classification accuracy for these classes. Towards addressing this issue, we observe that there exist natural correlations and anti-correlations among human-object interactions. In this paper, we model the correlations as action co-occurrence matri
A common problem in human-object interaction (HOI) detection task is that numerous HOI classes have only a small number of labeled examples, resulting in training sets with a long-tailed distribution. The lack of positive labels can lead to low classification accuracy for these classes. Towards addressing this issue, we observe that there exist natural correlations and anti-correlations among human-object interactions. In this paper, we model the correlations as action co-occurrence matrices and
Abstract A new type of processable electroluminescent polymer containing tertiary amine linkage was prepared by typical Wittig reaction between bis(4-formylphenyl) n-butylamine and diphosphonium salt. The resulting polymer was highly soluble in common organic solvents so that it could be spun-cast onto glass plate coated with ITO electrode to give highly transparent homogeneous thin film. The molecular weight of the polymer determined by gel permeation chromatography using polystyrene standards
The software industry has been increasingly aware of the need to employ copyright protection techniques against software piracy or illegal distribution. Recently, as the value of software as intellectual property grows, so does the rate of global software piracy. Software piracy is a real threat to software industry. Moreover, Pirated software is more vulnerable to attacks since such software is less likely to be supported with critical security patches and updates. In this paper, we propose a b
A software birthmark is unique, as certain native characteristics of a program, hence can be used to measure the similarity between programs. In general, a static software birthmark does not need program execution, but is more vulnerable to attacks by semantic-preserving transformations. A dynamic software birthmark is applicable to packed executables, but cannot cover all the possible program paths. In this paper, we propose a novel effective technique to measure the similarity of Microsoft Win
This paper addresses the problem in Web page ranking of effectively combining link and content information with efficiency high enough to be applicable to real-world search engines. Unlike previous surfer models, our approach is based on the viewpoint of a Web page author. Based on this viewpoint, we formulate the concept of contribution score, which indicates the amount to which a term in each page is utilized by other pages. To improve efficiency without loss of effectiveness, we exploit the e
Research Areas
Dive deeper into Dongjin Kim's research on Nubint
Open this lab's papers in the app to read with AI, summarize, and cite in your writing.