Lite Any Stereo V2: Faster and Stronger Efficient Zero-Shot Stereo Matching
Abstract Overview
Lite Any Stereo V2 (LAS2) is a family of efficient stereo matching models aimed at zero-shot generalization under practical deployment constraints. The paper redesigns the architecture around a purely 2D cost aggregation framework, replacing heavier 3D components to optimize measured latency on GPUs and edge devices rather than relying only on theoretical MACs. It also introduces a three-stage training pipeline that combines large-scale synthetic supervision, self-distillation, and real-world knowledge distillation from pseudo labels, with additional pseudo-label filtering and error clamping to stabilize synthetic-to-real transfer. The model family includes feed-forward variants (S/M/L) for different efficiency budgets and an iterative variant (H) for higher accuracy.
Novelty
The main novelty is the combination of a deployment-oriented 2D-only stereo architecture with a staged training recipe specifically designed for efficient zero-shot stereo matching. The work is also distinctive in how it improves real-world pseudo supervision through left-right consistency, edge and sky filtering, plus error-clamped optimization during adaptation.
Results
Across four zero-shot real-world benchmarks, LAS2-M consistently outperforms the prior feed-forward LAS model while running faster, and LAS2-H achieves stronger overall performance than Fast-FoundationStereo with substantially lower latency. Reported speedups include 1.6x/1.9x faster inference for LAS2-M over LAS on H200/Orin and 1.8x/2.7x faster inference for LAS2-H over Fast-FoundationStereo, while maintaining or improving benchmark accuracy. On the DrivingStereo weather split, LAS2 variants also show strong robustness, with LAS2-H achieving the best overall performance among efficient iterative methods.
Key Points
- LAS2 replaces hybrid or 3D-heavy aggregation with a purely 2D cost aggregation design chosen using measured deployment latency, improving the accuracy-latency trade-off on both server GPUs and edge devices.
- The training strategy is central to performance: synthetic pretraining, self-distillation, and real-world pseudo-label distillation are combined with pseudo-label filtering and error clamping to improve zero-shot transfer.
- Experiments show state-of-the-art accuracy among efficient stereo methods, faster inference than key feed-forward and iterative baselines, and competitive robustness on standard zero-shot and adverse-weather evaluations.