In June 2025, ICE CYPRESS, in collaboration with a leading university research team, published a landmark result in International Journal of Computer Vision (IJCV) — the premier journal in computer vision — titled “Implementation and Validation of Distributed Bundle Adjustment for Super Large Scale Datasets”
In traditional 3D reconstruction, the LM algorithm falls into a “memory explosion” trap due to ultra-large matrix operations. Existing workarounds resort to approximate solutions — sacrificing accuracy.
Put simply, the heart of 3D reconstruction is Bundle Adjustment (BA) — think of it as assigning precise “seats” to millions of cameras and 3D points so that their projected errors are minimized. But at scale, the traditional LM algorithm gets choked by a massive matrix called the Reduced Camera System (RCS / 改化法方程). For example, with 10 million images, the RCS becomes astronomically large!
Our solution gives this data a radical “Decluttering” treatment:
The distributed bundle adjustment (DBA) framework proposed by ICE CYPRESS parallelizes the formation of the reduced camera system (RCS) and incorporates the in-house block-based sparse matrix compression (BSMC) format, raising core-matrix storage efficiency by 1.5×–3×.
- Distributed Splitting — ICE CYPRESS’s distributed framework (DBA) parallelizes RCS generation across groups, letting each subgroup compute its own “sub-matrix” before assembling the global model. Like building with LEGO bricks — divide and conquer.
- The “Magic Compression” (BSMC) — We pioneered a Block-based Sparse Matrix Compression (BSMC) format, purpose-built for the sparse structure of RCS. Same data, 1.5–3× storage efficiency gain, with memory footprint slashed by up to 90%.
Crucially, our fully distributed BA achieves iterative optimization without ever aggregating a global matrix, breaking through the architectural scale ceiling of classical algorithms at the root level.
The 10M-image test reaches ~500× the scale of state-of-the-art LM-based BA pipelines. The technology is already integrated into commercial software Mirauge3D, powering city-cluster-level real-scene 3D modeling and UAV photogrammetric missions.
“Fully distributed algorithms make true planet-scale unified adjustment possible.”Through the co-design of distributed RCS formation and BSMC compression, the algorithm scales to exabyte-scale 3D reconstruction — delivering the critical technical backbone for smart cities, digital twins, and beyond.
