Nowadays we enjoy visual media based on technologies for image acquisition, compression, processing, transmission and displays. Recent researches deal with not only images but also light fields for advanced visual media. We introduce ray space representation of light fields and its various applications. For example, only with a single system of lenses, free viewpoint image reconstruction and scene refocusing with preferable bokeh are achieved. Moreover, acquired 3-D images allow us to observe scenes from arbitrary viewpoints simultaneously for 3-D consistent ambient communications, where dynamic light field reconstruction efficiently works for multi-view 3-D displays.Presenter : Kei Terasawa(NII Special Collaborative Research Student／Tokyo University of Science)

Event cameras offer high temporal resolution and low latency, making them effective for a range of applications. In this work, we explore their use in acquiring physical surface properties under active lighting. A key challenge is the absence of absolute intensity values. To overcome this, we design a custom setup with tailored equipment and temporally encoded lighting. Our method enables robust estimation of both surface normals and spectral reflectance. Results show that event cameras, when combined with designed illumination, can extract meaningful physical information, expanding their utility beyond motion detection and into the domain of material and shape analysis.Presenter : Art Subpaasa

Through light and computer-based analysis, we explore the secret behind the painting technique known as "milky white skin," brought to life by Léonard Foujita (Fujita Tsuguharu). While it had been difficult to determine how Foujita distinguished and applied various shades of white, our spectrofluorescence analysis has revealed the distribution of multiple white pigments.Presenter : Shin Ishihara

The scattering properties of biological tissue vary significantly depending on its health status and the presence of pathological changes.This poster introduces a technique for visualizing and evaluating tissue conditions using microscope images of tissue sections captured under structured illumination.Presenter : Yuta Asano

Noise canceling technology is used in earphones and ducts to cancel out noise using loudspeaker signals. However, when conventional technology is applied in three-dimensional space, the area that can be made quiet is very limited. We are conducting research on spatial active noise control based on the prediction and control of spatial sound to suppress noise over a large area. In this presentation, we will introduce our work on its practical application.Presenter : Denny Hermawanto, Shoichi Koyama

Fact-checking is like being a detective in the world of information, where the goal is to uncover the truth behind the statements or claims we hear and read every day in news articles, speeches, and social media posts. A fact-checker would also look into websites, medical studies, and research to find evidence for claims. Fact-checking is crucial in combating the rise of misinformation. Traditional methods of fact-checking involve human labour to verify claims through extensive research and cross-checking with credible sources, which is not only slow but also requires significant resources. Essentially, fact-checking has four main steps: identify claims, gather evidence, confirm the accuracy of facts, and explain the findings. In our study, we explore Automated Fact-Checking (AFC) through the use of Large Language Models (LLMs). We analyze over 17,000 real-world claims from PolitiFact.com, spanning from 2007 to 2024. We collect evidence of real-world claims from various websites. Then, we utilize a systematic approach to classify whether claims are true or false, and we also generate detailed explanations for each classification. We tested large models of 70 billion size to improve the accuracy of fact-checking and the quality of LLM’s explanations. The results show that bigger models performed better than traditional methods, even without training or fine-tuning of models. Additionally, incorporating evidence from various websites into the models improved LLMs accuracy in classifying real-world claims.Presenter : Maab Iffat

Nowadays, due to the burst of generative artificial intelligence, Deepfake voices become increasingly prevalent online, posing risks of misinformation and public confusion. In this poster you can listen and make guesses to both real and AI-generated short audio clips in both English and Japanese, and then see how our under-development deepfake detection system judges the same samples. Our detection system is based on a widely applicable 2B-parameter foundation model trained on 74,000 hours of natural and synthetic speech, including Japanese ones. For those interested, we are also happy to explain the machine learning technologies that power this deepfake detection system. We can also also explain the details of the machine learning technology underneath about deepfake detection and watermarking to those who are interested.Presenter : Xuechen Liu, Wanying Ge

Piano performance is more than playing the correct notes on a score: subtle changes in tempo and dynamics, the delicate color of each tone, and even the acoustics of the performance space combine to move listeners. Conventional AI methods have struggled to reproduce these human-like nuances under varied musical styles and recording conditions.
We introduce MIDI-VALLE, a next-generation AI piano synthesis system that leverages advanced language-modeling techniques. Although language models were originally developed to generate natural text, MIDI-VALLE treats both MIDI data and audio as “tokens” and learns from a wide range of styles and environments. As a result, it can generalize across genres—from classical to pop—and recreate expressive, lifelike performances even in unfamiliar acoustic settings.
In this poster, we will demonstrate how MIDI-VALLE applies language modeling to capture and reproduce the expressive qualities of human piano playing. Through audio examples, you will hear how the system preserves nuance—phrasing, dynamics, ambience—and opens new possibilities for AI-generated music. Presenter : Zhe Zhang, Yigitcan Oezer

The remarkable advancement of generative AI has increased the risk of misinformation spreading. Deepfakes are a type of misinformation that uses generative AI to artificially create content, such as videos or images, by replacing the faces of individuals with those of others, thereby causing harm such as the spread of false information. In this presentation, we will demonstrate how our developed deepfake detection system, SYNTHETIQ VISION, can identify deepfake videos as counterfeit. By leveraging SYNTHETIQ VISION's detection capabilities, which are capable of handling the latest generative AI technology, we invite you to experience the first step toward a future where we can address misinformation even in a society where AI has become widespread.Presenter : Yusuke Yasuda, Yoshihiko Furuhashi

NTCIR (NII Testbeds and Community for Information access Research）is a series of evaluation workshops of information access technologies including information retrieval, question answering, text generation, inference from text, etc.). It has worked together with more than 100 research groups from all over the world.Presenter : Noriko Kando, Makoto P. Kato

Search models based on large language models (vector search techniques) have been shown to outperform traditional search methods, and their adoption in real-world services is progressing. On the other hand, various issues have been revealed, such as cases where performance decreases compared to conventional technologies, unexpected behaviors occur, and certain search results are unintentionally favored. This poster introduces such "weaknesses" of vector search techniques.Presenter : Makoto P. Kato

Researchers of informatics need large scale data such as text, speech and video generated from sensors, social media, etc. We promote the shared use of such data by means of mediating between industrial bodies and researchers to stimulate research, deepen technology, and create communities. The posters will present the Center’s overview, summaries of the datasets provided through the “Informatics Research Data Repository (IDR)”, and some examples of the research using the datasets.Presenter : Tomoko Ohsuga