Pusan National University Researchers Develop Mixture of Experts Framework for Dynamic 3D Reconstruction

Researchers at Pusan National University have developed two mixture of experts frameworks for dynamic 3D scene reconstruction, the university announced in a press release. The work was made available online on July 13, 2026, in IEEE Transactions on Pattern Analysis and Machine Intelligence.

Dynamic 3D reconstruction models moving scenes over time, but different motion types can be difficult for one representation to capture. The team, led by Professor Kyeongbo Kong, designed MoE-GS and MoDE to combine multiple Dynamic Gaussian Splatting approaches instead of using a single motion model.

MoE-GS trains multiple dynamic Gaussian models separately and blends their outputs using learned expert routing. MoDE uses a shared Gaussian representation and integrates multiple deformation experts during joint optimization.

The researchers found that the combined expert approach improved reconstruction quality in scenes with several types of motion. The paper is titled On the Design of Mixture of Experts for Dynamic Gaussian Splatting.