StereoAdapter-2: Globally Structure-Consistent Underwater Stereo Depth Estimation
New AI model tackles underwater vision distortion with linear-time processing and a massive 80K-image synthetic dataset.
Researchers led by Zeyu Ren introduce StereoAdapter-2, a new AI model for underwater stereo depth estimation. It replaces ConvGRU with a ConvSS2D operator based on selective state space models (SSMs), enabling efficient long-range spatial propagation in a single step. The team also created UW-StereoDepth-80K, a large synthetic dataset. The system achieves state-of-the-art zero-shot performance with a 17% improvement on the TartanAir-UW benchmark, validated on real underwater robots like BlueROV2.
Why It Matters
Enables more reliable autonomous navigation and inspection for underwater drones and robotics in challenging, distorted environments.