---
license: mit
language:
- en
pretty_name: M³VIR
size_categories:
- 100B<n<1T
task_categories:
- image-to-3d
---
# M³VIR
In the field of restoration and 3D reconstruction, particularly for rendered content such as gaming environments, the lack of sufficient ground-truth training data presents a significant challenge. While these techniques are extensively studied in real-world applications, their adaptation to virtual or synthetic environments remains relatively underexplored. This is largely due to the distinct characteristics of rendered content, which differ from natural scenes in terms of texture, lighting, and geometry. The absence of high-quality, annotated datasets tailored for virtual content restoration and reconstruction has hindered progress and limited the development of effective methods in this domain. 

To address this gap, we introduce a large-scale, high-quality dataset specifically designed for rendered environments. This dataset aims to support a wide range of tasks, including image restoration, 3D reconstruction, novel view synthesis, and content manipulation, thereby facilitating research and development of generative AI algorithms for virtual content.

### Dataset Sources

- **Repository:** [M3VIR](https://huggingface.co/datasets/guluthemonster/M3VIR)
- **Paper:** [More Information Needed]

## Dataset Details

M³VIR provids 8 categories: Churches-And-Temples, Hiking-Trails, Hotels-And-Restaurants, Mountains, Parks-And-Recreation-Areas, Residential-Areas, School-Universities, and Urban-Street-Views. For each category, we collected three types of scenes: 
- MovingCameraDynamicScene
- MovingCameraStaticScene
- StaticCameraDynamicScene

For each scene type, we collect 10 distinct video sets featuring varying scene content. Each set includes different resolutions and visual styles: a photo-realistic style available in 960×540, 1920×1080, and 2880×1620 resolutions (Realistic_960x540_1024sample, Realistic_1920x1080_1024sample, Realistic_2880x1620_1024sample); a cartoon style in 1920×1080 resolution (Cartoon_1920x1080_1024sample); and a metalized style also in 1920×1080 resolution (Metalize_1920x1080_1024sample). Corresponding segmentation maps are provided as ID_images. Since Realistic_1920x1080_1024sample, Cartoon_1920x1080_1024sample, and Metalize_1920x1080_1024sample share the same segmentation annotations, we include the ID_images only once to conserve storage.

The dataset is split into 80% for training (64 sets) and 20% for testing (16 sets). To support the four challenge tracks, the full M³VIR dataset is divided into two subsets: M³VIR_MR and M³VIR_MS. Due to the large size of the dataset, a small-scale mini training set will also be provided for Track 1 to facilitate quick experimentation and baseline development.

Each video sample—defined by a specific style and resolution (e.g., realistic style at 1920×1080 resolution)—includes six temporally synchronized camera views with a shared camera center. These views are captured from different perspectives: Back, Front, Left60, Left120, Right60, and Right120, providing diverse angular coverage of the scene. Each video sequence is 2 seconds long, recorded at 15 frames per second, resulting in a total of 30 image frames per view.

### M³VIR-Tracks
| Dataset           | Rate|Scene Types|Resolution Styles|Data Path|
| :-----------      | :-----------:       |  :-------:       |  :-------:       |  :-------:       | 
| M³VIR_MR     | 5%    |MovingCamDyn/MovingCamStatic/StaticCamDyn |Real_960x540/Real_1920x1080/Real_2880x1620|Track1|
|              | Full  |MovingCamStatic|Real_1920x1080|Track2|
|              | Full  |MovingCamStatic|Real_960x540/Real_1920x1080/Real_2880x1620|Track3|
| M³VIR_MS     | Full  |MovingCamDyn/MovingCamStatic/StaticCamDyn |Cartoon_1920x1080/Metal_1920x1080/Real_1920x1080|Track4|


For more details about the datasets and challenge tracks, please refer to the official challenge page:
https://richmediagai.github.io/challenges.html

## How to Download
Our dataset is divided into a training set (TRACKS) and a test set (TEST), each serving different purposes. For more details, please refer to train_list.txt and test_list.txt.

[Use Hugging Face Command Line Interface (CLI)](https://huggingface.co/docs/huggingface_hub/guides/cli#huggingface-cli-download)

```
Download Entire Dataset

$ huggingface-cli download guluthemonster/M3VIR --repo-type dataset --local-dir .

Download Specified Folder

$ huggingface-cli download guluthemonster/M3VIR --repo-type dataset --include TRACKS/* --local-dir .
```

[Use Git](https://huggingface.co/docs/hub/datasets-downloading#using-git)
```
$ git clone https://huggingface.co/datasets/guluthemonster/M3VIR
```

After download the dataset, you can use following codes to extract the files in each subfolder (take the TRACKS/Track1/Batch1 as an example)：

```
$ python Scripts/extract_track1.py --input_path TRACKS/Track1/Batch1 --output_path /path/to/your/folder
```
# Citation