XinhuiLiu001/M3ISR

Categories

Category    | Scenes
------------|------------------------------------
Bedroom     | Bedroom_01, Bedroom_02, Bedroom_03, Bedroom_04, Bedroom_05
Kitchen     | Kitchen_01, Kitchen_02, Kitchen_03, Kitchen_04, Kitchen_05
LivingRoom  | Livingroom_01, Livingroom_02, Livingroom_03, Livingroom_04, Livingroom_05
Outdoor1    | Outdoor_01, Outdoor_02, Outdoor_03, Outdoor_04, Outdoor_05
Outdoor2    | Outdoor_06, Outdoor_07, Outdoor_08, Outdoor_09, Outdoor_10
Total: 25 Scenes

M³ISR

In the field of 3D scene reconstruction and novel view synthesis, access to high-quality multi-view data covering both indoor and outdoor environments remains limited. Real-world datasets face constraints such as capture cost, lighting variability, and privacy concerns, while most existing synthetic datasets are either indoor-only or outdoor-only and rarely provide temporally synchronized multi-view captures with co-registered multi-modal annotations. The absence of a unified synthetic dataset that spans everyday indoor–outdoor settings, together with dense depth and semantic ground truth, has hindered progress on reconstruction and generative algorithms for residential and outdoor virtual content.

To address this gap, we introduce M³ISR, a synthetic dataset featuring multi-view, multi-modal captures across multiple indoor and outdoor scene categories. M³ISR is designed as a companion to M³VIR, extending the rendered-scene paradigm from landmark and outdoor-focused environments to everyday indoor–outdoor settings. The dataset aims to support a wide range of tasks, including 3D reconstruction, novel view synthesis, dynamic scene modeling, and multi-modal learning, thereby facilitating research and development on reconstruction and generative AI algorithms for virtual content.

Dataset Sources

Repository: M3ISR

Paper: [More Information Needed]

Dataset Details

M³ISR provides 5 categories covering both indoor and outdoor environments: Bedroom, Kitchen, LivingRoom, Outdoor1, and Outdoor2. For each category, we collected two types of scenes:

• MovingCameraDynamicScene
• MovingCameraStaticScene

For each category, we provide 5 distinct scene instances featuring varying scene content (e.g., Bedroom_01 to Bedroom_05), resulting in 25 scenes in total. All scenes are rendered in a photo-realistic style at 1920×1080 resolution with 1024 samples per pixel (Realistic_1920x1080_1024sample). Corresponding segmentation maps are provided as ID_images, and ground-truth depth maps are provided as Depth_images where available. A per-scene ColorsSummary.xlsx documents the mapping between semantic IDs and their RGB colors.

Unlike conventional scene-level splits, M³ISR adopts a frame-level hold-out scheme widely used in novel view synthesis benchmarks (e.g., the llffhold=N convention in NeRF and 3D Gaussian Splatting). Within each view of each scene, every 6-th frame is reserved as a test frame while the remaining frames are used for training, yielding an exact 5:1 train-to-test ratio per view. This rule is applied independently to each view angle and to both MovingCameraDynamicScene and MovingCameraStaticScene subsets, ensuring that every scene contributes to both training and evaluation. A reference implementation is provided in Scripts/split_frames.py.

Each scene sample—defined by a specific scene type and rendering configuration (e.g., moving-camera static scene at 1920×1080 photo-realistic style)—includes six temporally synchronized camera views with a shared camera center. These views are rendered simultaneously at six horizontal rotation angles around the camera's vertical axis: −60°, −40°, −20°, +20°, +40°, and +60°, 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, 180 frames per scene sample across the six views, and 9,000 frames across the full dataset.

M³ISR-Summary

Quantity	Value
Categories	5
Scenes per category	5
Total scenes	25
Scene types per scene	2
Synchronized views per scene type	6
Frames per view (2s @ 15fps)	30
Frames per scene sample (all views)	180
Total RGB frames (full dataset)	9,000
Training frames (full dataset)	7,500
Test frames (full dataset)	1,500

How to Download

Our dataset is organized under the repository root, grouped by category. Each scene is stored as a self-contained subfolder so that users may download and work with individual scenes or categories without retrieving the entire dataset. The train/test split is defined at the frame level and can be generated on the fly using Scripts/split_frames.py.

Use Hugging Face Command Line Interface (CLI)

Download Entire Dataset

$ huggingface-cli download XinhuiLiu001/M3ISR --repo-type dataset --local-dir .

Download a Single Category

$ huggingface-cli download XinhuiLiu001/M3ISR --repo-type dataset --include "Bedroom/*" --local-dir .

Download a Single Scene

$ huggingface-cli download XinhuiLiu001/M3ISR --repo-type dataset --include "Bedroom/Bedroom_01/*" --local-dir .

Use Git

$ git clone https://huggingface.co/datasets/XinhuiLiu001/M3ISR

After downloading the dataset, you can generate the frame-level train/test split using the following script:

$ python Scripts/split_frames.py --dataset_root . --output_dir splits --hold_out 6

This produces splits/train.txt and splits/test.txt, each listing the paths of the corresponding frames.

License

This dataset is released under the MIT License.

Acknowledgement

M³ISR is inspired by and complementary to the M³VIR dataset (https://huggingface.co/datasets/guluthemonster/M3VIR). We thank the M³VIR authors for establishing the rendered-scene dataset paradigm that M³ISR follows.