Fixstars Open-Sources CUMVS: Multi-View Stereo Software for Rapid 3D Model Generation from 2D Images

Fixstars Corporation, a global leader in AI-powered software development and acceleration, today announced the open-source release of its Multi-View Stereo (MVS) software, “CUMVS (cuda-multi-view-stereo),” under the Apache License 2.0. CUMVS is designed to accelerate the realization of digital twins of the physical world.

Also Read: How are Collaborative Data Ecosystems Driving AI Adoption in Regulated Industries?

3D Model Reconstruction method and Multi-View Stereo

3D model reconstruction technology provides a means to recreate objects existing in the physical world within a virtual 3D space. This technology is used in a variety of applications, including the construction of digital twins in cyber-physical systems, the digital preservation of cultural heritage, simulations for architecture and urban planning, and the design of customized medical devices tailored to patient needs.

Multi-View Stereo (MVS) is an essential technique for reconstructing 3D models of subjects from multiple images. It plays a crucial role in dense point cloud reconstruction and texture reproduction. Fixstars’ CUMVS significantly improves the processing speed of MVS, enabling the rapid creation of high-definition 3D models.

CUMVS: Optimized for NVIDIA GPUs

Fixstars has successfully developed CUMVS from scratch, optimizing the PatchMatch MVS algorithm, a key algorithm in MVS, for NVIDIA GPUs. Leveraging Fixstars’ extensive optimization expertise, CUMVS incorporates NVIDIA GPU-specific optimizations such as preventing Warp Divergence [1], utilizing shared memory and texture units, and streamlining heavy processes like Homography transformation through approximation formulas. CUMVS achieves a speed increase of more than 5x compared to ACMM [2], a leading implementation of PatchMatch MVS.

Also Read:  AiThority Interview with Eli Ben-Joseph, CEO at Regard

This speed boost empowers researchers to process more complex models faster and allows developers to build innovative applications more efficiently.

[1] Warp Divergence: A situation in which individual processing cores within a GPU diverge in different directions. This branching often leads to variations in processing time for each core, generally reducing the efficiency of GPU utilization.

[2] Measurement Conditions:
Benchmark: ETH3D High-res multi-view training dataset
Measurement Machine: Intel Core i7-13700K, NVIDIA GeForce RTX 3080
Processing Time: ACMM approx. 12277 seconds, CUMVS approx. 2140 seconds (5.73x faster)

[To share your insights with us as part of editorial or sponsored content, please write to psen@itechseries.com]