Fast-FoundationStereo: Real-Time Zero-Shot Stereo Matching
Fast-FoundationStereo accelerates the three bottlenecks of FoundationStereo separately: distillation compresses the hybrid monocular+stereo backbone into one efficient student, blockwise NAS searches cost-filtering designs under latency budgets with ILP, and structured pruning guided by a recurrent dependency graph slims the ConvGRU refinement module. Trained with 1.4M auto pseudo-labeled in-the-wild stereo pairs, it runs over 10x faster than FoundationStereo while nearly matching zero-shot accuracy, setting a new state of the art among real-time methods on Middlebury, ETH3D, KITTI and Booster.
Bowen Wen, Shaurya Dewan, Stan BirchfieldDec 11, 2025
Stereo MatchingDepth EstimationNVIDIADec 11, 2025
