A Joint Research Team Led by Professor Kang Suk-ju of the Department of Electronic Engineering Has Paper Accepted at ‘ECCV 2026,’ a Top Conference in the Field of Artificial Intelligence
A joint research team consisting of Professor Kang Suk-ju’s group (Yeom Su-woong and Nam Jun-sik, Integrated Master’s-Ph.D. program; Choi Seung-kyu, Master’s program) from the Department of Electronic Engineering at this university, Dr. Kim Jun-soo and Dr. Yoon Kook-jin (Electronics and Telecommunications Research Institute), Lucas Yunkyu Lee (Pohang University of Science and Technology), Kim Sang-min and Professor Park Jae-sik (Seoul National University), and Professor Kong Kyeong-bo (Pusan National University) has had a paper officially accepted for presentation at “ECCV 2026 (European Conference on Computer Vision),” a most prestigious international conference in the fields of computer vision and artificial intelligence. Together with CVPR and ICCV, both organized by IEEE/CVF, ECCV is regarded as one of the world’s top three conferences in computer vision and machine learning.
The title of the paper is “TRiGS: Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting.” In this work, the research team proposed the “TRiGS” model, a new 4D Gaussian Splatting (4DGS) technique capable of effectively rendering long-duration dynamic videos, for the first time.
The team pointed out that existing 4D Gaussian Splatting models suffer from a problem known as “temporal fragmentation” when handling complex object motions. This occurs because these models rely on short time windows or piecewise linear velocity approximations, forcing them to constantly remove and regenerate Gaussian splats. As a result, the long-term temporal identity of objects is compromised, and memory usage skyrockets, posing serious limitations for application to long video sequences.
To address these performance bottlenecks, the team moved away from piecewise approximation methods and developed TRiGS, which optimizes rigid-body motion by utilizing “continuous geometric transformations.”
The core of this research lies in the “combined SE(3) transformation,” which integrates complex rotational and translational motions into a single continuous transformation, and the “Hierarchical Bezier Residuals” block, which precisely captures nonlinear motion. Furthermore, the team achieved stable and precise local motion optimization by introducing “Local Anchors,” which assign independent rotation centers to each Gaussian rather than relying on a single global center.
As a result, by fundamentally suppressing unnecessary Gaussian proliferation, the model achieved results that surpass those of existing state-of-the-art models while limiting the total number of model parameters to approximately 0.5 million (around 160 MB). Experimental results demonstrate that TRiGS performs overwhelmingly stable dynamic view synthesis on major benchmark datasets such as N3V and SelfCap, even for long sequences ranging from 600 to 1,200 frames, without significant memory overhead or degradation in rendering quality.
The students who participated in this research expressed their thoughts, stating, “We were able to achieve these excellent results thanks to Professor Kang Suk-ju’s thoughtful guidance and the generous support of our lab colleagues. We are very pleased that our paper has been accepted by a top-tier computer vision conference, and we will continue to strive to produce meaningful research results in the future.”
▶Paper Title: TRiGS: Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting
▶Author Information: Yeom Su-woong (Integrated Master’s-Ph.D. Student, Co-first author), Nam Jun-sik (Integrated Master’s-Ph.D. Student, Co-first author), Choi Seung-kyu (Master’s Student, Co-first author), Lucas Yunkyu Lee, Kim Sang-min, Professor Park Jae-sik, Dr. Kim Jun-soo, Dr. Yoon Kook-jin, Professor Kong Kyeong-bo (Co-corresponding author), and Professor Kang Suk-ju (Co-corresponding author)
[SEO keyword]
4D Gaussian Splatting, Dynamic View Synthesis, TRiGS Model
The title of the paper is “TRiGS: Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting.” In this work, the research team proposed the “TRiGS” model, a new 4D Gaussian Splatting (4DGS) technique capable of effectively rendering long-duration dynamic videos, for the first time.
The team pointed out that existing 4D Gaussian Splatting models suffer from a problem known as “temporal fragmentation” when handling complex object motions. This occurs because these models rely on short time windows or piecewise linear velocity approximations, forcing them to constantly remove and regenerate Gaussian splats. As a result, the long-term temporal identity of objects is compromised, and memory usage skyrockets, posing serious limitations for application to long video sequences.
To address these performance bottlenecks, the team moved away from piecewise approximation methods and developed TRiGS, which optimizes rigid-body motion by utilizing “continuous geometric transformations.”
The core of this research lies in the “combined SE(3) transformation,” which integrates complex rotational and translational motions into a single continuous transformation, and the “Hierarchical Bezier Residuals” block, which precisely captures nonlinear motion. Furthermore, the team achieved stable and precise local motion optimization by introducing “Local Anchors,” which assign independent rotation centers to each Gaussian rather than relying on a single global center.
As a result, by fundamentally suppressing unnecessary Gaussian proliferation, the model achieved results that surpass those of existing state-of-the-art models while limiting the total number of model parameters to approximately 0.5 million (around 160 MB). Experimental results demonstrate that TRiGS performs overwhelmingly stable dynamic view synthesis on major benchmark datasets such as N3V and SelfCap, even for long sequences ranging from 600 to 1,200 frames, without significant memory overhead or degradation in rendering quality.
The students who participated in this research expressed their thoughts, stating, “We were able to achieve these excellent results thanks to Professor Kang Suk-ju’s thoughtful guidance and the generous support of our lab colleagues. We are very pleased that our paper has been accepted by a top-tier computer vision conference, and we will continue to strive to produce meaningful research results in the future.”
▶Paper Title: TRiGS: Temporal Rigid-Body Motion for Scalable 4D Gaussian Splatting
▶Author Information: Yeom Su-woong (Integrated Master’s-Ph.D. Student, Co-first author), Nam Jun-sik (Integrated Master’s-Ph.D. Student, Co-first author), Choi Seung-kyu (Master’s Student, Co-first author), Lucas Yunkyu Lee, Kim Sang-min, Professor Park Jae-sik, Dr. Kim Jun-soo, Dr. Yoon Kook-jin, Professor Kong Kyeong-bo (Co-corresponding author), and Professor Kang Suk-ju (Co-corresponding author)
[SEO keyword]
4D Gaussian Splatting, Dynamic View Synthesis, TRiGS Model