{"id":26698,"date":"2026-06-06T00:00:00","date_gmt":"2026-06-05T22:00:00","guid":{"rendered":"https:\/\/saarland-informatics-campus.de\/?post_type=sic_news&p=26698"},"modified":"2026-06-09T10:39:39","modified_gmt":"2026-06-09T08:39:39","slug":"dfki-at-cvpr-2026-from-guided-3d-scenes-to-spinal-simulation","status":"publish","type":"sic_news","link":"https:\/\/saarland-informatics-campus.de\/en\/piece-of-news\/dfki-at-cvpr-2026-from-guided-3d-scenes-to-spinal-simulation\/","title":{"rendered":"DFKI at CVPR 2026: From guided 3D scenes to spinal simulation"},"content":{"rendered":"

With several papers presented at CVPR 2026, DFKI demonstrated the breadth of its research in visual AI. The spectrum ranged from 3D scene understanding and relational reasoning, through multimodal perception, to simulation, generation and workshop contributions.<\/strong><\/a><\/p>\n

The IEEE\/CVF Conference on Computer Vision and Pattern Recognition, or CVPR for short, is one of the most important conferences in the field of computer vision research and took place this year from 3 to 7 June in Denver. The DFKI was represented there with several accepted papers from various areas of research. The focus was on a paper from the Augmented Vision research group that addresses a key weakness in current 3D scene analysis: systems recognise objects but often fail to understand how they relate to one another.<\/p>\n

<\/a><\/p>\n

\n

\n Main conference papers from the DFKI
\n <\/h2>\n<\/header>\n

The DFKI contributions to the main conference cover a broad spectrum of visual AI. From the Augmented Vision research area come \u201cReLaGS: Relational Language Gaussian Splatting\u201d, \u201cDriverGaze360: OmniDirectional Driver Attention with Object-Level Guidance\u201d, \u201cLiREC-Net: A Target-Free and Learning-Based Network for LiDAR, RGB, and Event Calibration\u201d, and \u201cSIMSPINE: A Biomechanics-Aware Simulation Framework for 3D Spine Motion Annotation and Benchmarking\u201d.\u00a0<\/p>\n

These are joined by \u201cOpenMarcie: A dataset for multimodal action recognition in industrial environments\u201d and \u201cWhen Pretty Isn\u2019t Useful: An investigation into why modern text-to-image models fail as reliable generators of training data\u201d, as well as \u201cYieldSAT: A multimodal benchmark dataset for high-resolution yield prediction\u201d from Kaiserslautern, \u201cSynthesising Visual Concepts as Vision-Language Programs\u201d from Darmstadt, and \u201cSceMoS: Scene-Aware 3D Human Motion Synthesis by Planning with Geometry-Grounded Tokens\u201d from Saarbr\u00fccken. Collectively, the topics range from open 3D scene understanding and multimodal perception to sensor calibration, medical simulation, synthetic training data and generative motion modelling.<\/p>\n

<\/a><\/p>\n

\n

\n ReLaGS
\n <\/h2>\n<\/header>\n

Within this spectrum, ReLaGS stands out. The paper by Yaxu Xie, Abdalla Arafa, Alireza Javanmardi, Christen Millerdurai, Jia Cheng Hu, Shaoxiang Wang, Alain Pagani and Didier Stricker combines a hierarchical 3D scene representation with an explicit scene graph that models relationships between objects. This makes it possible not only to identify objects within a scene, but also to process relational queries such as \u2018the cup next to the laptop\u2019 or more nuanced part-whole relationships within complex 3D environments.<\/p>\n

The method is based on Gaussian splatting, a state-of-the-art technique for high-resolution 3D reconstruction. ReLaGS supplements this with linguistic semantics and relational reasoning, organises scenes hierarchically \u2013 from parts through objects to the entire space \u2013 and does not require scene-specific training.<\/p>\n

<\/a><\/p>\n

<\/div>\n
\n

\u201cWith ReLaGS, we have shown that 3D scene understanding need not stop at the recognition of individual objects. The key lies in modelling relationships, hierarchies and semantic contexts together \u2013 only then does reconstruction truly become machine understanding.\u201d<\/p>\n