XStereo is a large-scale, high-fidelity synthetic stereo matching dataset rendered with Unreal Engine 5. It contains about 200K stereo pairs with dense disparity GT, collected from 162 carefully selected indoor and outdoor scenes.

XStereo is intended for:

Pre-training stereo matching / depth estimation models, then fine-tuning or zero-shot evaluating on real datasets.
Studying data scaling laws for stereo matching: in-domain error, saturation behavior, data redundancy, and effective data utilization.
Benchmarking the synthetic-to-real gap and the effect of scene/sample diversity on zero-shot generalization.
Data-efficient training research: nested subsets of 2.5K → 5K → 10K → ... stereo pairs can be drawn from the full set.

Highlights

Physically realistic rendering. UE5 pipeline with shadows, reflections, gloss, randomized lighting, lens flare and environmental fog.
Curated scenes, not random assets. 162 real indoor/outdoor projects with rich texture and fine geometric detail, rather than randomly placed and textured objects.
Human-like camera motion. The virtual camera walks inside the traversable area of each scene, producing viewpoints close to real-world captures.
Quality controlled. Up to 2,000 samples per scene, followed by human quality assurance that removes over-exposed and texture-less ill-posed frames to reduce label ambiguity.
Scalable by construction. Nested sub-datasets of increasing size can be sampled without introducing distribution shift, which makes XStereo a clean benchmark for data scaling research.

Dataset at a Glance

Property	Value
Total stereo pairs	~200K
Scenes	162 (indoor + outdoor)
Renderer	Unreal Engine 5
Image resolution	512 × 960 (H × W)
Focal length	500–1100 px
Stereo baseline	0.06–0.25 m
Modalities	Rectified left image, rectified right image, disparity GT

Getting Started

PyTorch dataloader example:

import os
from PIL import Image
import numpy as np
import torch.utils.data as torch_data
class XStereoDataset(torch_data.Dataset):
    def __init__(self, datalist_file, max_disp):
        super().__init__()
        
        self.disp_range = (0.01, max_disp)
        self.root = 'dataset dir'
        self.data_list = []
        with open(datalist_file, 'r') as fp:
            self.data_list.extend([x.strip().split(' ') for x in fp.readlines()])
    
    def __len__(self):
        return len(self.data_list)
    
    def read_disp(self, left_depth_path):
        img = Image.open(left_depth_path)  
        color_array = np.asarray(img, dtype=np.uint8)  
        r = color_array[..., 0].astype(np.uint32)
        g = color_array[..., 1].astype(np.uint32)
        b = color_array[..., 2].astype(np.uint32)
        
        depth_uint32 = (r << 16) | (g << 8) | b
        depth = depth_uint32.astype(np.float32) / 10 # in cm
        splits = left_depth_path.split('_')
        f, b = int(splits[-2]), int(splits[-1].split('.')[0])   # focal in pixel, baseline in cm

        # for some scenes, the depth values larger than 500 meters are denoted as 0
        # for these cases, dividing a small epsilon resulting a huge disparity, which is then truncated by max_disp
        return f*b / (depth + 1e-6) 
    
    def __getitem__(self, idx):
        item = self.data_list[idx]  
        full_paths = [os.path.join(self.root, x) for x in item]
        left_path, right_path, left_depth_path = full_paths

        # image, 960*512
        left_img = Image.open(left_path).convert('RGB')
        left_img = np.array(left_img, dtype=np.float32)
        right_img = Image.open(right_path).convert('RGB')
        right_img = np.array(right_img, dtype=np.float32)

        # disp
        disparity = self.read_disp(left_depth_path)
        sample = {
            'left': left_img,
            'right': right_img,
            'disp': disparity,
        }

        sample['valid'] = (self.disp_range[0] < sample['disp']) & (sample['disp'] <= self.disp_range[1])
        
        return sample

Third-Party Content Notice

The 3D scenes in this dataset were rendered using Unreal Engine 5 and may incorporate third-party assets obtained from various sources, including Fab Marketplace .

All third-party assets remain the intellectual property of their respective rights holders and are subject to their original license terms.

The dataset authors do not claim ownership of the underlying third-party assets. Any applicable attribution requirements, usage restrictions, and redistribution conditions associated with these assets must be respected by users.

The assets are only used for producing the dataset for research purposes. Neither the dataset nor the assets are related to any commercial activities within XPENG INC..

The dataset authors' licensing terms apply only to the original contributions made by the authors, including dataset organization, annotations, metadata, and associated code, to the extent that these materials are independently copyrightable.

This dataset is provided for research purposes. Users are responsible for ensuring that their use of the dataset complies with applicable licenses and laws.

License

The released files (left/right images, disparity maps and any accompanying annotations) are distributed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

You are free to share and adapt the data for non-commercial purposes, provided that appropriate credit is given. Commercial use is not permitted.

License deed: https://creativecommons.org/licenses/by-nc/4.0/
Full legal code: https://creativecommons.org/licenses/by-nc/4.0/legalcode

Citation

@article{zhao2026stereoscaling,
  title  = {Data Scaling Laws in Stereo Matching: From In-domain Fitting to Cross-domain Generalization},
  author = {Zhao, Tong and Liang, Di and Zeng, Yiming and Xu, Bo and Tang, Yingzhi and Bai, Chen},
  year   = {2026},
  note   = {arXiv:2610.38600},
  url    = {https://arxiv.org/abs/2610.38600}
}

Contact

If you have any questions, concerns about the dataset or the assets, please contact the dataset authors at zhaot9@xiaopeng.com.

Downloads last month: 20

Total file size:

19.5 MB

Paper for Tong-Zhao/XStereo

Data Scaling Laws in Stereo Matching: From In-domain Fitting to Cross-domain Generalization

Paper • 2610.38600 • Published 7 days ago