MVImgNet: A Large-scale Dataset of Multi-view Images

by Xianggang Yu*, Mutian Xu*†, Yidan Zhang*, Haolin Liu*, Chongjie Ye*, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, Guanying Chen, Shuguang Cui, Xiaoguang Han‡ from GAP-Lab.

📢 News

• ❗️We renew the download script to fix (Jul. 8, 2025 UTC):
  • The SharePoint authentication may interrupt the download process.
  • Users can now choose the specific folder to download the MVImgNet.
• 💪 MVImgNet2.0 is released! It expands MVImgNet to a total of ~520k real-life objects and 515 categories, and contains ∼300k real-world objects in 340+ classes. The annotation comprehensively covers object masks, camera parameters, and point clouds (Feb. 20, 2025 UTC).

Introduction

This repository is built for:

MVImgNet: A Large-scale Dataset of Multi-view Images (CVPR2023) [arXiv]

If you find our work useful in your research, please consider citing:

@inproceedings{yu2023mvimgnet,
    title     = {MVImgNet: A Large-scale Dataset of Multi-view Images},
    author    = {Yu, Xianggang and Xu, Mutian and Zhang, Yidan and Liu, Haolin and Ye, Chongjie and Wu, Yushuang and Yan, Zizheng and Liang, Tianyou and Chen, Guanying and Cui, Shuguang and Han, Xiaoguang},
    booktitle = {CVPR},
    year      = {2023}
}
}

MVImgNet

MVImgNet contains 6.5 million frames from 219,188 videos crossing objects from 238 classes. We provide a SharePoint link to download the full data. Please fill out this form to get the download link and password.

Usage 🌟🌟🌟

We provide two ways to download MVImgNet:

1. Directly download from SharePoint web interface. Then:

cd path/to/mvimgnet_zip_file
unzip './*.zip'

2. We also provide the script at download_tool.zip. Usage:

(Before using it, please make sure you have filled out our form and have received the password)

• Unzip the download_tool.zip using the password offered in form. The password provided in the form can be used to download data from the SharePoint web interface as well as to unzip the download tool.
• Choose the subfolder to download (e.g., [MVImgNet_origin, MVImgNet_category, MVImgNet_mask])
• Download with (e.g., using MVImgNet_origin):

python download_tool.py --data_name "MVImgNet_origin" --download_folder "/path/you/want/to/save"

🚩 NOTE:

• We split the full data into 43 zip files (from mvi_00.zip to mvi_42.zip), and the total size is about 3.4 TB. If you need the full dataset, please directly download all mvi_xx.zip.
• If you just need some categories of the data, please download the corresponding categories at MVImgNet_by_categories, where we split the dataset into different categories (but it is incomplete, some categories are not included).

Folder structure

|-- ROOT
    |-- class_label
        |-- instance_id
            |-- images
            |-- sparse/0
                |-- cameras.bin   # COLMAP reconstructed cameras
                |-- images.bin    # binary data of input images
                |-- points3D.bin  # COLMAP reconstructed sparse point cloud (not dense) 

The mapping between class_label and class name can be found in mvimgnet_category.txt.

The images folder contains the multi-view images, and the sparse folder contains the reconstructed camera parameters using COLMAP. It is recommended to use the functions provided by COLMAP to read the binary files under sparse folder. Moreover, the gen_poses function from this repo is recommended to convert the poses for NeRF training.

MVPNet

MVPNet now contains 87,825 point clouds from 180 categories. Please fill out the following form to download MVPNet.

Demo

MVImgNet is also shown by Voxel51 at here, which is publicly demo-able!

License

The data is released under the MVImgNet Terms of Use, and the code is released under the Attribution-NonCommercial 4.0 International License.

Copyright (c) 2023

Acknowledgement

Thanks to Wei Cheng for the new downloading solution for our dataset.

Thanks to Gege Gao for providing tips on downloading our dataset.

Thanks to Voxel51 for providing the dataset demo.