Reproduce “Lite Any Stereo V2: Faster and Stronger Efficient Zero-Shot Stereo Matching” on OpenResearch

About this paper

Junpeng Jing, Ronglai Zuo, Zhelun Shen, Shangchen Zhou, Rolandos Alexandros Potamias, Stefanos Zafeiriou, Krystian Mikolajczyk, Jiankang Deng

Imperial College London researchers developed Lite Any Stereo V2 (LAS2), a series of ultra-fast stereo matching models designed for efficient zero-shot generalization across diverse real-world scenarios. LAS2-M demonstrates up to 1.9 times faster inference on NVIDIA Orin with improved accuracy over prior efficient methods, while the LAS2-H iterative variant surpasses high-capacity models like FoundationStereo in zero-shot tasks while being 2.7 times faster.

Key insights

• High efficiency and strong zero-shot generalization can be achieved concurrently within lightweight stereo models through targeted architectural and training innovations.
• Optimizing architectures for practical deployment latency, rather than solely theoretical FLOPs, is crucial for real-world performance on modern hardware.
• Robust pseudo-label filtering and error-clamping are critical for stabilizing the synthetic-to-real transfer process and effectively leveraging unlabeled real-world data for generalization.

Summary by alphaXiv.