FUSE-Flow enables real-time 3D reconstruction without calibration targets

Real-time multi-camera 3D reconstruction is critical for immersive media and spatial computing, but existing systems suffer from tightly coupled calibration and fusion steps that introduce cumulative errors and poor scalability. A new paper on arXiv introduces FUSE-Flow, a decoupled framework that splits the pipeline into two collaborative modules: GMAC for geometry-aligned multi-view extrinsic calibration and FUSE for reliability-guided stateless point cloud fusion. This split design allows each module to be optimized independently, eliminating conflicting objectives and enabling more robust performance.

The GMAC module refines camera extrinsics using geometric constraints and multi-view reconstruction transformers, achieving accurate calibration from sparse views without requiring calibration targets, dense images, or global bundle adjustment. The FUSE module integrates confidence weighting and adaptive spatial hashing to perform stateless fusion with linear time and memory complexity. The two modules reinforce each other — accurate poses improve fusion quality, while confidence-aware fusion corrects calibration biases. Validated on public datasets and real camera setups, FUSE-Flow outperforms mainstream real-time reconstruction methods in visual effect, dynamic stability, and scalability, offering a practical solution for large-scale applications.