Video by Meta Open Source via YouTube
The Project Aria team is officially releasing code for Aria Gen 2 Stereo Depth Estimation with NVIDIA’s FoundationStereo. This includes a tutorial of the full steps to perform stereo rectification and depth-from-stereo, as well as a tool to export dense metric depth from any Aria Gen 2 recording.
Our depth-from-stereo pipeline is built on top of two critical components: a highly accurate and generalizable stereo foundation model, FoundationStereo, and pixel-accurate online calibration of the stereo cameras on Aria Gen 2 devices with Visual Inertial Odometry (VIO). FoundationStereo is trained on over 1 million photorealistic synthetic pairs and delivers accurate stereo disparity estimates on Aria Gen 2 data without any fine-tuning needed. To convert these disparity estimates into metric depth estimates we leverage our state-of-the-art VIO system running on-device to provide continuously refined stereo calibration parameters as the glasses move and change over time.
Why this matters:
🤖 Zero-Shot Generalization: Works in diverse environments without retraining.
⚡️ High Performance: Real-time ready, capable of 20 FPS (with TensorRT).
🛠️ Full Pipeline: Covers stereo image rectification and disparity estimation.
🔓 100% Open Source: Built for the research and developer community.
Whether you are building for Robotics, AR/VR, or SLAM, Project Aria Gen 2’s depth-from stereo capability provides the robust depth foundation you’ve been waiting for.
👉 GitHub Tutorial & Tools: https://github.com/facebookresearch/projectaria_gen2_depth_from_stereo
#ProjectAria #ComputerVision #OpenSource #Robotics #MachineLearning #AITools #StereoVision