News

Youngrae attended ECCV 2026 in Malmö, Sweden, where he presented two posters from MCL. The first was Xinyu Wang’s PASR, which detects camouflaged objects by identifying regions that deviate from common scene patterns. The second was his own work, MemRoPE, a training-free method that enables consistent long-video generation using a fixed-size memory.

As long-video generation is currently a popular research topic, the MemRoPE poster attracted many researchers and industry representatives. They were particularly interested in how the method can be applied to existing video-generation models without additional training and how it maintains consistent subjects and scenes over long durations.

Yann LeCun also delivered a keynote in which he strongly emphasized his vision for JEPA-based world models. Rather than predicting every pixel, JEPA predicts missing or future information in an abstract representation space. The broader goal is to help AI develop an internal model of the world that supports understanding, prediction, and planning.

Overall, ECCV was a valuable opportunity to present our work, receive direct feedback from the research community, and learn about recent developments in video generation and world models.

We are very happy to welcome a new MCL member, Wen-Yuan Ting. Here is a quick interview with Wen-Yuan:

Could you briefly introduce yourself and your research interests?My name is Wen-Yuan Ting. I received my B.S. degree in Electrical Engineering from National Tsing Hua University and my M.S. degree in Communication Engineering from National Taiwan University. After receiving my M.S. degree, I was a Research Assistant at Academia Sinica for over two years, specializing in machine learning. My current research interests are wireless communications, machine learning, and signal processing

What is your impression of MCL and USC? MCL is a very warm lab, and I feel like we have a great academic atmosphere. Professor Kuo is a very supportive mentor, and I am deeply honored to be one of his Ph.D. students. As for USC, my first impression is that USC has a very beautiful campus, with students coming from many backgrounds. I am also impressed by the strong faculty at USC, which provides many opportunities to learn from world-class experts.

What is your future expectation and plan in MCL?Since my current research topic focuses on the intersection of wireless sensing and communications, I plan to publish papers in IEEE Transactions journals, preferably in venues such as IEEE Transactions on Wireless Communications (TWC) and IEEE Transactions on Signal Processing (TSP). As I have always been interested in interpreting theoretical mathematics from an engineering perspective, I hope that my publications will demonstrate my mathematical understanding and provide interpretable engineering designs to the world.

SPIE 2026 Optics + Photonics Exhibition, held at the San Diego Convention Center, was conducted from 8/23 to 8/27 over five days. The conference featured both poster and oral presentations, alongside an extensive exhibition floor where numerous companies showcased optics- and photonics-related technologies spanning applications from imaging systems and sensors to laser and semiconductor manufacturing. I presented my paper, “Medical Image Classification: Methodology and Performance Evaluation on MedMNIST,” as part of the medical imaging session, which featured three oral presentations that afternoon. It is composed of a 15-minute presentation and a 5-minute Q&A. During the Q&A for my talk, an audience member raised a question regarding the computational aspects of my method. One of the other two presentations in the session were also from our MCL which is EEG connectivity maps classification by Qi Cao. Overall, the conference offered valuable opportunities to engage with the broader optics and photonics community, receive constructive feedback on our work, and gain insight into how classification methodologies are being applied across diverse imaging modalities.

Gaussian Splatting has recently emerged as a powerful representation for reconstructing real-world scenes from images, offering fast optimization and real-time, high-quality novel-view rendering. However, its original design is primarily rendering-oriented: a collection of Gaussian primitives does not directly provide the explicit surface geometry required by many downstream applications such as simulation, animation, AR/VR, and conventional graphics pipelines.

Recent research is therefore shifting toward geometry-aware Gaussian Splatting, where Gaussians are equipped with more meaningful geometric information such as depth, surface normals, and continuous occupancy or vacancy fields. These developments have enabled increasingly accurate mesh extraction directly from Gaussian representations. For example, recent methods can reconstruct detailed surfaces, including thin structures, while retaining the efficient rendering properties of Gaussian Splatting.

Our research investigates the next step in this direction: budget-controlled GS-to-mesh compilation. Instead of first extracting a very dense mesh and simplifying it afterward, we aim to use the geometry and rendering information already stored in the Gaussian representation to directly allocate mesh elements where they are most useful. Given a desired polygon budget, the goal is to efficiently produce compact mesh assets at different levels of detail while preserving geometric and visual quality.

The Amazon Prime collaboration focused on the detection of artifacts and distortions present in streaming applications. Specifically, we aim to detect both transient and persistent artifacts within professional video feeds with long time horizons and no reference. Previous approaches to artifact detection lack the computational efficiency and generalizability to identify the wide range of potential distortions that may occur while monitoring thousands of streams at scale.

Our solution extracts features across disjoint domains combining standard natural scene statistics with novel ones including Laws’ filter responses. These features are then scored against a pristine reference distribution to characterize deviations from normal and pooled across frames in a short temporal window. Standard green learning tools, including the DFT and LNT, are utilized to identify and generate additional highly discriminative features for an XGBoost classifier. The green learning tools are organized into a multi-round framework, so that each successive round iteratively improves upon previous predictions for difficult samples. Experiments demonstrate that our method is both parameter efficient and accurate with performance exceeding comparable deep learning methods. In parallel, we developed a dataset of short and long form videos, based on LongVideoBench, with various streaming-specific distortions to aid further research in long-form professional video artifact detection.