sanjay810/mv2-dataset

MV2 Dataset

MV2 is a multi-view and multi-vehicle urban driving dataset designed for research in novel view synthesis, neural rendering, 3D reconstruction, cross-view scene understanding, and autonomous-driving perception. The dataset contains synchronized image sequences captured from multiple viewpoints, including ground vehicles and aerial views, along with camera parameters required for geometry-aware learning and rendering.

The dataset is released for academic and research use. The full dataset is being made available here, and additional annotations such as segmentation masks for dynamic objects may be released in upcoming updates.

Dataset Details

Dataset Description

MV2 focuses on challenging real-world urban driving scenes with large viewpoint changes across vehicles and platforms. It is intended to support research on:

• Cross-view novel view synthesis
• Multi-vehicle and multi-camera scene reconstruction
• Ground-to-aerial and aerial-to-ground rendering
• Geometry-aware neural rendering
• Autonomous-driving scene understanding
• Dynamic-scene analysis in urban traffic environments

Unlike single-ego-vehicle driving datasets, MV2 includes multiple observation platforms, making it useful for studying extreme-baseline view synthesis and shared world modeling.

Dataset Sources

• Repository: TODO: add Hugging Face dataset URL
• Project page: TODO: add MV2 project page URL
• Paper: TODO: add paper/arXiv link
• Code: TODO: add GitHub repository link

Dataset Structure

The dataset is organized at the sequence level. Each sequence contains multiple views/platforms such as car, drone, scooty, and additional car-view variants.

A typical structure is:

MV2/
├── seq1/
│   ├── car/
│   │   ├── images/
│   │   ├── poses/
│   │   └── intrinsics/
│   ├── drone/
│   │   ├── images/
│   │   ├── poses/
│   │   └── intrinsics/
│   ├── scooty/
│   │   ├── images/
│   │   ├── poses/
│   │   └── intrinsics/
│   └── car_front_left_forward/
│       ├── images/
│       ├── poses/
│       └── intrinsics/
├── seq2/
│   └── ...
└── ...

Each view contains RGB frames and corresponding camera parameters. File names follow a sequence-based frame indexing format such as:

car_front_000000.jpg
car_front_000000.npy
drone_000000.jpg
drone_000000.npy
scooty_front_000000.jpg
scooty_front_000000.npy

Data Fields

The dataset may contain the following fields depending on the sequence and view:

Field	Description
images/	RGB image frames for each camera/view
poses/	Camera pose matrices stored as .npy files
intrinsics/	Camera intrinsic matrices stored as .npy files
scale_factors/	Scale information for aligning reconstructed geometry, if provided
masks/	Dynamic object segmentation masks, planned for release in later updates

Views / Platforms

The dataset includes multiple viewpoints:

View	Description
car	Ground vehicle camera sequence
scooty	Two-wheeler/scooter-mounted camera sequence
drone	Aerial camera sequence
car_front_left_forward	Additional car-view variant for cross-view evaluation

Usage

You can download the dataset using the Hugging Face Hub:

from huggingface_hub import snapshot_download

dataset_path = snapshot_download(
    repo_id="YOUR_USERNAME/MV2",
    repo_type="dataset"
)

print(dataset_path)

Example for loading an image and camera parameters:

from pathlib import Path
from PIL import Image
import numpy as np

root = Path(dataset_path)

image_path = root / "seq1" / "drone" / "images" / "drone_000000.jpg"
pose_path = root / "seq1" / "drone" / "poses" / "drone_000000.npy"
intrinsics_path = root / "seq1" / "drone" / "intrinsics" / "drone_000000.npy"

image = Image.open(image_path).convert("RGB")
pose = np.load(pose_path)
intrinsics = np.load(intrinsics_path)

print(image.size)
print("Pose:", pose.shape)
print("Intrinsics:", intrinsics.shape)

Intended Uses

MV2 is intended for academic and research applications, including:

• Novel view synthesis
• Neural rendering
• Multi-view 3D reconstruction
• Cross-view image synthesis
• Aerial-to-ground and ground-to-aerial scene understanding
• Dynamic urban-scene modeling
• Autonomous-driving perception research

Out-of-Scope Uses

The dataset should not be used as the sole source for deploying safety-critical autonomous-driving systems. It is a research dataset and may not cover all environmental, geographic, weather, lighting, or traffic conditions required for real-world deployment.

Dataset Creation

The dataset was collected in real-world urban driving environments using multiple camera platforms. The goal was to capture diverse viewpoints of the same or related scenes, enabling research on cross-view consistency, geometry-aware rendering, and shared scene understanding.

Camera parameters are provided to support 3D reconstruction, novel view synthesis, and geometry-based evaluation.

Ethical Considerations

The dataset contains real-world urban driving imagery. Users should avoid using the data for identifying individuals, tracking private persons, or any surveillance-related application. Researchers should follow applicable privacy, data protection, and responsible AI guidelines when using the dataset.

Contact

For questions, issues, or access-related queries, please contact:

Sanjay Bhargav Dharavath
sanjaytinku810@gmail.com

Updates

• Initial full dataset release.
• Dynamic object segmentation masks are planned for a future update.