FlowObject: Flow Steering for Bridging Generative Priors and Reconstruction Fidelity
Researchers have introduced FlowObject, a novel framework designed to improve 3D object reconstruction from limited image captures. This method addresses the limitations of existing generative models, which can exhibit "synthetic bias" and disregard observational data, and optimization-based methods that struggle with unseen geometry. FlowObject employs a dual-space guidance strategy to steer the trajectory of flow-matching models, allowing for the completion of occluded regions using generative priors while maintaining consistency with real-world observations. A subsequent 3D Gaussian Splatting refinement stage further enhances photorealism and observational fidelity, outperforming current state-of-the-art methods in both geometric completeness and appearance. AI
IMPACT This framework could significantly improve the accuracy and completeness of 3D models generated from sparse image data, impacting fields like AR/VR and digital content creation.