Meta releases XRStereo dataset for 3D scene reconstruction from stereo video
Meta has published the XRStereo synthetic dataset on Hugging Face to train models for reconstructing 3D scenes from stereo video.

Meta has released a new dataset called XRStereo on the Hugging Face platform, designed to train computer vision models and systems. This synthetic dataset is built specifically for the task of reconstructing three-dimensional scenes using standard stereo video inputs. The total volume of the data is nearly 380 gigabytes, providing researchers in spatial computing with a powerful resource.
- Contains nearly 18,000 sequences with five stereo pairs in each.
- The resolution of every individual image is 1024 by 1024 pixels.
- Includes 13 virtual environments, featuring 9 indoor and 4 outdoor scenes.
- The dataset archive size is approximately 373 GB, split into 72 separate archives.
Structure and parameters of the dataset
Within each sequence, virtual cameras move through a static scene. Along with standard left and right images, the creators included precise data on camera positions, the baseline distance between lenses, and disparity maps representing pixel shifts between the two perspectives. This allows algorithms to deeply analyze spatial depth and train neural networks to better understand 3D environments.
During the generation of these virtual environments, researchers varied shooting parameters such as the horizontal field of view, ranging from 50 to 130 degrees, and the distance between cameras, set from 4 to 12 centimeters. Such diversity enables the dataset to be used for training algorithms tailored to various stereo cameras found in modern VR headsets and robotics.
Read alsoNvidia confirms open-source AI strategy and Hugging Face acquisitionConnection to the StereoSplat method and accessibility
The primary purpose of creating XRStereo was to support the StereoSplat method, developed by Meta engineers alongside specialists from the University of Amsterdam. This approach builds 3D representations using Gaussian Splatting and then generates novel views to show how a scene would look if the camera position shifted.
Alongside the dataset itself, Meta released the source code for the StereoSplat method and the pre-trained model weights. The entire resource is hosted on Hugging Face and divided into 72 individual archives totaling nearly 373 gigabytes. This openness is designed to accelerate research and development in computer vision and 3D graphics.
Significance for developers and the industry
The introduction of such open datasets greatly simplifies the work of engineers and scientists developing virtual and augmented reality technologies. The ability to train neural networks on high-quality synthetic data with finely tuned camera parameters allows for the creation of more realistic immersive products. Open-source code and ready-to-use model weights give developers a strong foundation for spatial video projects.
Source: Все публикации подряд на Хабре



