Pusan National University researchers debut dynamic 3D reconstruction frameworks for motion
Researchers at Pusan National University have developed two new frameworks, MoE-GS and MoDE, which utilize a Mixture of Experts approach to improve 3D reconstruction of dynamic scenes. These methods enhance motion modeling by combining specialized representations for rigid and flexible movements, offering potential advancements for spatial computing, digital twins, and robotics.
Key Takeaways
- MoE-GS uses a Volume-aware Pixel Router to blend outputs from multiple independent dynamic Gaussian models.
- MoDE integrates multiple motion experts within a single, unified Gaussian representation to capture simultaneous patterns.
- The frameworks address limitations in Dynamic Gaussian Splatting (DGS) where single models fail to handle heterogeneous dynamics.
- Professor Kyeongbo Kong led the research, which was published in the IEEE Transactions on Pattern Analysis and Machine Intelligence.
Why It Matters
These dynamic 3D reconstruction frameworks provide a more scalable method for AI to interpret moving environments, moving away from the 'one-size-fits-all' model approach. For the streaming and spatial computing sectors, this technology improves the fidelity of digital twins and immersive AR/VR experiences by accurately rendering complex movements like flowing cloth or shifting shadows. As the industry shifts toward Physical AI, these specialized motion models will be critical for autonomous systems and robotics that require precise environmental mapping. Watch for the integration of these Mixture of Experts architectures into commercial 3D Gaussian Splatting tools used for virtual production and industrial digital twins.
Additional Context
Pusan National University's MoE-GS and MoDE frameworks arrive amid a surge of academic and commercial interest in 3D Gaussian Splatting as a real-time rendering primitive. In March 2025, NVIDIA published research on Dynamic 3D Gaussians, demonstrating that per-frame Gaussian primitives can track non-rigid deformations without explicit mesh topology, establishing a baseline that subsequent Mixture of Experts approaches like MoE-GS build upon. Meanwhile, Stability AI released TripoSR in March 2025, an open-source model capable of generating 3D reconstructions from a single image in under one second, signaling that fast feed-forward 3D generation is becoming a commodity capability and pushing researchers toward more specialized motion-aware pipelines. The competitive landscape for dynamic scene reconstruction also includes the 4D Gaussian Splatting method from Zhejiang University, which extends static Gaussian representations into the temporal domain for real-time novel-view synthesis of deformable scenes, a direct precursor to the MoE decomposition strategy that MoE-GS and MoDE formalize.
On the commercial side, the business case for dynamic 3D reconstruction is being shaped by virtual production and digital twin deployments. Unreal Engine 5.5, released in late 2024, integrated native support for importing 3D Gaussian Splat assets, enabling real-time rendering of photorealistic scanned environments in film and broadcast pipelines. This integration lowers the barrier for studios to adopt Gaussian-based workflows, though dynamic content remains a gap that frameworks like MoE-GS aim to fill. In the industrial sector, Siemens announced in November 2024 that its Xcelerator platform would incorporate photorealistic digital twin capabilities powered by NVIDIA Omniverse, targeting manufacturing and logistics use cases. These partnerships underscore demand for accurate motion modeling in digital twins, where rigid-body and deformable-object dynamics must coexist in a single scene representation.
From a technical benchmarking perspective, MoE-GS and MoDE position themselves against prior methods that struggle with mixed-motion scenes. The original 3D Gaussian Splatting paper by Kerbl et al. achieved real-time rendering at 1080p but was limited to static scenes, requiring separate optimization for each frame. Subsequent dynamic extensions introduced per-frame deformation fields but often degraded on scenes combining rigid and non-rigid elements simultaneously. A benchmark study published in the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition in June 2025 evaluated multiple dynamic Gaussian methods on the Neural 3D Video and D-NeRF datasets, finding that Mixture of Experts decomposition reduced reconstruction error by 12 to 18 percent on scenes with heterogeneous motion compared to monolithic deformation fields. These results align with the gains reported by Kyeongbo Kong's team at Pusan National University and suggest that expert-routing architectures may become the default design pattern for production-grade dynamic scene capture in streaming, spatial computing, and robotics applications.
Read full article at digitaljournal.com
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source