Dronescapes
Drone videos with human-annotated semantic segmentation, SfM depth and surface normals, recorded over several scenes in Romania and Norway.
RGB video, 8-class semantics, SfM depth, Surface normals, Optical flow, Edges
Datasets
Datasets released with our papers: aerial video, multi-modal drone recordings, synthetic 3D scenes and Earth observation layers. We are moving all of them to Hugging Face, where each can be downloaded with a singlegit clone.
On Hugging Face
Dronescapes Experts
Dronescapes with extra modalities produced by pretrained expert models through the Video Representations Extractor (VRE).
RGB video, Expert-generated modalities, Semantics, Depth, Normals
NEO: NASA Earth Observations, monthly
Monthly global NASA Earth Observations layers at 540×1080, prepared for multi-task learning, with loaders, viewer notebooks and train/test splits.
AOD, Chlorophyll, Cloud fraction, NDVI, Temperature, + more
Not yet on Hugging Face
These are still on SharePoint, Google Drive or MEGA. The links below work until they move.
SafeUAV synthetic aerial 3D
Synthetic aerial images rendered from 3D reconstructions of real places, with depth and safe-landing labels.
Synthetic RGB, Depth, Plane orientation / safe-landing labels
Ruralscapes
4K aerial videos of rural Romania with dense manual labels every 50 frames. When we released it, it was the largest aerial video dataset for semantic segmentation.
4K aerial video, Dense semantic labels every 50 frames, 12 classes
UAV metric depth (UFO Depth / Depth Distillation)
Drone flights with synchronized telemetry, used to learn metric depth without depth ground truth (ICRA 2022, CVPR 2021 Workshops).
Drone video, Kinematics / odometry, Optical flow, Metric depth
Neural Graph Consensus multi-task data
Multi-task scene data used in the Neural Graph Consensus experiments (AAAI 2021). The code is at gitlab.com/mihaicristianpirvu/cycleconcepts.
RGB, Depth, Semantics, Normals, Camera pose
European Buildings
Satellite images of European cities with building masks.
Satellite RGB, Building masks
European Roads
Satellite images of European cities with road masks.
Satellite RGB, Road masks
Aerial geolocalization (roads & intersections)
Aerial images with road and intersection annotations, for locating an image without GPS (BMVC 2016).
Aerial RGB, Road / intersection labels, GPS
AnoShift
Anomaly detection benchmark with distribution shift, built from ten years of network traffic (NeurIPS 2022 Datasets and Benchmarks).
Network traffic (Kyoto-2006+), 10+ years, Temporal splits
OpenLLM-Ro benchmarks & SFT data
Romanian versions of standard LLM benchmarks and instruction-tuning sets, released with the OpenLLM-Ro models.
Romanian text, Instruction tuning, Benchmarks (ARC, MMLU, GSM8K, …)
Details
Dronescapes
Released with Self-supervised Hypergraphs for Learning Multiple World Interpretations (ICCV 2023 Workshops). It has train, validation, semi-supervised and test splits, and the dataset card includes the evaluation scripts and a benchmark table. To download: git lfs install && git clone https://huggingface.co/datasets/Meehai/dronescapes.
Dronescapes Experts
Everything except the human-annotated semantics and the SfM depth and normals can be regenerated from the raw videos with VRE.
NEO: NASA Earth Observations, monthly
Includes the scripts to download the raw data, convert it from PNG to NPY and build the splits. A weekly version is planned.
SafeUAV synthetic aerial 3D
The Hugging Face repository exists but only has a README so far. Download the data from SharePoint for now.
Ruralscapes
Sixteen student volunteers from Petru Cercel High School in Târgoviște helped label it. It was used with SegProp for label propagation and semi-supervised segmentation.
UAV metric depth (UFO Depth / Depth Distillation)
Neural Graph Consensus multi-task data
European Buildings
European Roads
Aerial geolocalization (roads & intersections)
AnoShift
Maintained by Bitdefender’s machine learning group (bit-ml).