StereoFactory: A Unified Merging Framework for Robust Stereo Matching
Researchers have developed StereoFactory, a novel framework for merging specialized stereo matching models into a more robust system. This approach uses a two-stage evolutionary process, first employing a genetic algorithm to select optimal model subsets and then refining architecture-adaptive routing with CMA-ES optimization. Experiments show StereoFactory significantly reduces error rates on benchmarks like NMRF and FoundationStereo compared to baseline methods, while requiring a fraction of the computational time of joint retraining. AI
IMPACT Introduces a more efficient method for combining specialized AI models, potentially reducing training costs and improving performance in computer vision tasks.