Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

Meta* Releases XRStereo: Nearly 18,000 Stereo Frame Sets for 3D Vision

VR headsets and robots need thousands of left and right camera frame pairs to see a room in three dimensions, and capturing them in the real world is expensive. Meta* has released XRStereo, a synthetic dataset of nearly 18,000 sets, each containing five stereo pairs at 1024 × 1024 pixels, drawn from 13 virtual scenes. It comes with the code for the StereoSplat method and ready-made weights, so the model can be tested on your own recordings without training from scratch.

AI-processed from Habr AI; edited by Hamidun News
Meta* Releases XRStereo: Nearly 18,000 Stereo Frame Sets for 3D Vision
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The dataset weighs about 373 GB and is split into 72 archives on Hugging Face
Key metric #2The horizontal field of view ranges from 50° to 130°, and the distance between cameras from 4 to 12 cm
Key metric #3The scenes include nine interiors and four outdoor settings, with cameras moving through a static scene

For a VR headset or a robot to understand the room around it, a flat image is not enough. It needs a volumetric model. That model can be built from two frames captured by left and right cameras, the same way our eyes do it. Training such systems requires thousands of examples where the camera positions are known exactly, and capturing them in the real world takes a long time and costs a lot. Meta has published the XRStereo dataset on Hugging Face: nearly 18,000 sets (Meta calls them sequences), each containing five stereo pairs. That is roughly 90,000 frame pairs, all rendered in virtual scenes.

What's inside the dataset

XRStereo is a synthetic dataset. Because the scenes are created in virtual environments, the annotations are exact and require no manual labeling. Each image has a resolution of 1024 × 1024 pixels. In each sequence, the cameras move through a static scene: only the viewpoint changes, while the world itself stays put.

In addition to the left and right frames, the data includes camera poses and the distance between the cameras. It also includes disparity maps. To put it simply: if you look at a finger first with one eye and then with the other, it "jumps" relative to the background. The closer the object, the bigger the jump. A disparity map records the amount of this shift for every pixel, and from it you can calculate the distance to each object.

There are 13 environments: nine interiors and four outdoor scenes. The number of locations is modest, but the variety comes from the capture parameters.

Why the camera angle and baseline were varied

Stereo cameras differ from device to device. One headset has its lenses closer together, another has them farther apart, and a robot's cameras may sit on a separate bar altogether. If a model is trained on one geometry, it usually performs worse on another.

To reduce this dependence, the researchers varied the horizontal field of view from 50° to 130° and the distance between cameras from 4 to 12 cm when creating the scenes. By their design, this allows systems to be trained for the stereo cameras of different VR headsets and robots, rather than for one specific device.

StereoSplat: the purpose behind the dataset

The dataset was created primarily for the StereoSplat method, which Meta* developed together with specialists from the University of Amsterdam. The method builds a three-dimensional representation of a scene using Gaussian Splatting. In this approach, a scene is described by a large number of soft colored "blobs" in space, from which an image can be assembled quickly. The system then generates an image from new viewpoints, showing how the room would look if the camera were moved to the side.

In practice, this means that a volumetric model can be obtained from video captured by a stereo camera, rather than a set of flat frames. This is useful in VR and AR, where the picture must be re-rendered to match the position of the head. It is also useful in robotics, where a machine needs to estimate the layout of the space around it.

Availability and limitations

XRStereo takes up nearly 373 GB and is split into 72 archives. For a small team, that is a substantial volume: it requires disk space and download time. Alongside the data, Meta* has released the StereoSplat code and weights for pre-trained models. This means you do not have to train everything from scratch. You can first run the ready-made model on your own stereo recordings and assess its quality.

There are also clear limitations. The dataset is synthetic, and the gap between rendered and real-world capture is typically significant in computer vision: sensor noise, glare, and motion blur look different in virtual scenes. All sequences were captured in static scenes, so moving people and objects are not represented in the data. The source also does not indicate how well StereoSplat performs on real devices: no such results are included in the announcement.

The practical takeaway: if you are building VR or AR applications, or vision systems for robots with a stereo camera, XRStereo is a ready-made foundation for training and comparing methods. Check whether your camera's geometry falls within the 50–130° field-of-view range and the 4–12 cm baseline range, and only then decide whether synthetic data is sufficient or whether you will need to add real-world data.

*Meta Platforms, as well as its social networks Facebook and Instagram: Meta is recognized as an extremist organization, and its activities are banned in Russia; Facebook and Instagram are banned in Russia.

Story

Key Questions Answered

What's inside the dataset?

XRStereo is a synthetic dataset. Because the scenes are created in virtual environments, the annotations are exact and require no manual labeling. Each image has a resolution of 1024 × 1024 pixels. In each sequence, the cameras move through a static scene: only the viewpoint changes, while the world itself stays put. In addition to the left and right frames, the data includes camera poses and the distance between the cameras. It also includes disparity maps.

Why the camera angle and baseline were varied?

Stereo cameras differ from device to device. One headset has its lenses closer together, another has them farther apart, and a robot's cameras may sit on a separate bar altogether. If a model is trained on one geometry, it usually performs worse on another. To reduce this dependence, the researchers varied the horizontal field of view from 50° to 130° and the distance between cameras from 4 to 12 cm when creating the scenes.

What are the key details on StereoSplat: the purpose behind the dataset?

The dataset was created primarily for the StereoSplat method, which Meta developed together with specialists from the University of Amsterdam. The method builds a three-dimensional representation of a scene using Gaussian Splatting. In this approach, a scene is described by a large number of soft colored "blobs" in space, from which an image can be assembled quickly. The system then generates an image from new viewpoints, showing how the room would look if the camera were moved to the side.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…