ClaraVid Dataset

Official repo for: ClaraVid: A Holistic Scene Reconstruction Benchmark From Aerial Perspective With Delentropy-Based Complexity Profiling - Accepted ICCV 2025

If you find this useful, please consider giving us a like ❤

ClaraVid is a synthetic dataset built for semantic and geometric neural reconstruction from low altitude UAV/aerial imagery. It contains 16,917 multimodal frames collected across 8 UAV missions over diverse environments: urban, urban high, rural, highway, and nature. Each mission features 3 viewpoints and altitude levels, simulating multi-UAV operations. The dataset spans 1.8km^2, with an average mission coverage of 0.22km^2. It includes visual measurements at 4032x3024 resolution for RGB images, metric depth maps, panoptic(semantic and instance) segmentation and dynamic object masks. Additionally in contains scene level pointcloud and camera calibration(intrinsic and extrinsic).

Channel Log / TODOs

All data uploaded
Release dataset SDK
Release pip package
Release dataset splits
Add Nerfstudio support
Dataset download script
Release DSP code (closer to conference)

3. Download

Clone this repository and extract the data archive in the same folders. The archives were compressed using 7-Zip. We will provide a download script in the near future.

4. Usage

We provide a dataset SDK on GitHub. You can simply install it using:

pip install claravid

5. Dataset structure

All collection missions follow a grid pattern with both vertical and horizontal passes at a constant altitude, with a few seconds between consecutive frames.

claravid/
├── 001_rural_1/               # mission 1
│   ├── left_rgb/              
│   │   ├── 45deg_low_h/       # 3 different viewpoints (pitch&altitude) & flying orientation in grid (horizontal or vertical passes) 
│   │   │   ├── 000000.jpg
│   │   │   └── ...
│   │   ├── 45deg_low_v/
│   │   │   └── ...
│   │   ├── 55deg_mid_h/
│   │   │   └── ...
│   │   ├── 55deg_mid_v/
│   │   │   └── ...
│   │   ├── 90deg_high_h/
│   │   │   └── ...
│   │   └── 90deg_high_v/
│   │       └── ...
│   ├── depth/                 # metric depth
│   │   └── ...
│   ├── panoptic_seg/          # instance (buildings, humans and vehicles) & semantic segmentation
│   │   └── ...
│   ├── semantics_colormap/    # semantic segmentation - RGB color version
│   │   └── ...
│   ├── dynamic_mask/           
│   │   └── ...
│   ├── extrinsics/            
│   │   └── ...
│   └── scene_pcl/             # scene level PCL (color, semantic, instance) @ various resolutions (30cm, 50cm, ...)
│       ├── panoptic_seg/
│       │   ├── global_fused_30cm.ply
│       └── ...
├── 002_rural_2/               # mission 2...
│   └── ...
└── ...

6. Mission Overview

6. Data format

Modality	Directory	Extension	Description
RGB	left_rgb	.jpg	4032 x 3024
Depth	depth	.png	metric depth - [0-1000]m
Panoptic Segmentation	panoptic_seg	.png	instance (buildings, humans and vehicles) + semantic mask
Dynamic mask	dynamic_mask	.png	binary mask for objects that move (dynamic_elements == 0)
Camera Extrinsics	extrinsics	.json	in scene space (metric)
ScenePointcloud	scene_pcl	.ply	scene pointclouds in scene space

Please refer to our SDK for more details regarding the data format.

7. BibTex

If you find our work useful, please consider citing:

@article{beche2025claravid,
  title={ClaraVid: A Holistic Scene Reconstruction Benchmark From Aerial Perspective With Delentropy-Based Complexity Profiling},
  author={Beche, Radu and Nedevschi, Sergiu},
  journal={arXiv preprint arXiv:2503.17856},
  year={2025}
}

Datasets:

radubeche
/

claravid

Access Request for ClaraVid Dataset