By submitting this request, you confirm that you have read and agree to the following terms:
• You have reviewed and accept the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/)
• You will properly acknowledge and cite the dataset in any derivative works or publications

UAVid++: Higher-Quality Labels and Expanded Semantic Taxonomy for Aerial Semantic Segmentation

Vivian Chiciudean, Sergiu Nedevschi, Florin Oniga — Department of Computer Science, Technical University of Cluj-Napoca, Romania

Dataset Summary

High-quality annotated datasets are critical for semantic segmentation, particularly in aerial imagery, where the number of manually annotated frames is limited and precise labeling is often missing. To address this, we introduce UAVid++, a semantically refined and expanded version of the UAVid dataset for aerial semantic segmentation. The proposed dataset provides higher-quality annotations through manual corrections addressing labeling errors across the entire original UAVid dataset. In addition, we expand the taxonomy with three finely annotated semantic classes, Water, Sky, and Roof, selected to enrich the semantic space and separate visually distinct objects previously grouped together under the same label.

Variant	Description
UAVid	The original dataset (8 classes), used as reference. Can be downloaded at uavid.nl.
UAVid+	Corrected annotations only - labeling errors and contours fixed across the entire dataset, without the expanded taxonomy (8 classes).
UAVid++	Corrected annotations and the expanded semantic taxonomy with three new classes (11 classes).

Supported Tasks

Semantic segmentation of low-altitude urban UAV imagery. Evaluation metric: per-class Intersection-over-Union (IoU) and mean IoU (mIoU).

Dataset Structure

UAVid++ follows the original UAVid organization.

After downloading and extracting the dataset, the directory structure should look as follows:

.
├── uavid_train/
│   ├── seq1/
│   │   ├── Images/
│   │   │   ├── 000000.png
│   │   │   ├── ...
│   │   └── Labels/
│   │       ├── 000000.png
│   │       ├── ...
│   ├── seq2/
│   └── ...
├── uavid_val/
│   ├── seq16/
│   └── ...
└── uavid_test/
    ├── seq21/
    └── ...

where (IMPORTANT):

Images/ contains the original 4K RGB frames, identical to those in the original UAVid dataset. We recommend downloading them from the official UAVid release whenever possible. If that is not possible, you can use the UAVid_rgb_only.zip archive provided in this repository.
Labels/ contains the pixel-level semantic segmentation masks corresponding to each RGB image. The label files depend on the annotation version you choose: download the appropriate *_label.zip archive for either the UAVid++, UAVid+, or original UAVid annotations.

Each seq* directory corresponds to a single UAV video sequence.

uavid_train/, uavid_val/, and uavid_test/ are the official training, validation, and test splits.

For additional preprocessing scripts use the code repository.

Data Splits

The official UAVid split is preserved for direct comparability:

Split	Frames
Train	200
Validation	70
Test	150
Total	420

Semantic Taxonomy

UAVid++ extends the original 8-class UAVid taxonomy with three new classes - Water, Sky, and Roof. Segmentation masks use the following fixed class → color mapping (identical to the official UAVid colors for the original 8 classes).

ID	Class	RGB (hex)	RGB (decimal)	Notes
1	Background Clutter	`#000000`	(0, 0, 0)
2	Building (Wall)	`#800000`	(128, 0, 0)
3	Road	`#804080`	(128, 64, 128)
4	Tree	`#008000`	(0, 128, 0)
5	Low Vegetation	`#808000`	(128, 128, 0)
6	Dynamic Car	`#400080`	(64, 0, 128)
7	Static Car	`#C000C0`	(192, 0, 192)
8	Human	`#404000`	(64, 64, 0)
9	Water	`#0000FF`	(0, 0, 255)	New
10	Sky	`#80FFFF`	(128, 255, 255)	New
11	Roof	`#464646`	(70, 70, 70)	New

Water - includes all visible water surfaces, encompassing the lakes, ponds, swimming pools, and puddles present in the dataset, regardless of reflectance or mirroring. Objects floating on or partially submerged in water retain their original class labels down to the waterline. Wet ground without pooled water is not labeled as Water.
Sky - covers all visible sky regions, including clouds and haze. Backlit scene entities retain their foreground label up to the visible outline.
Roof - covers all upper surfaces of a building, namely pitched, mansard, flat, and stepped roofs, as well as roof-mounted equipment, such as the air-conditioning units and solar panels present in the dataset.

Experimental Results

State-of-the-art methods and the proposed approach, evaluated on the three variants (mIoU %, train/test on the matching variant):

Method	UAVid	UAVid+	UAVid++
UNetFormer	67.87	72.83	75.07
LWGANet	69.18	73.52	77.66
SFA-Net	69.52	75.85	78.43
LSKNet	69.03	76.44	78.61
D2LS	69.68	76.96	79.35
Ours (DINOv3 ViT-H+, UNet-inspired head)	72.23	78.67	81.42

Full per-class tables, ablations, and out-of-distribution / cross-domain results are present in the paper.

To reproduce the results reported in the paper, use the code repository.

Model Weights

output_vits_unet_uavid++.zip is the pretrained checkpoint for the DINOv3 ViT-S + UNet configuration (smoke test).
output_vith+_unet_uavid++.zip is the pretrained checkpoint for the DINOv3 ViT-H+ + UNet configuration.
output_vit7b_unet_uavid++.zip is the pretrained checkpoint for the DINOv3 ViT-7B + UNet configuration.

These checkpoints are optional and are only required if you want to run inference without training the model from scratch.

Intended Uses & Attribution

UAVid++ is intended for research on aerial/UAV semantic segmentation, annotation-quality studies, and generalization analysis. It is a high-precision benchmark (420 frames) focused on low-altitude oblique urban scenes, not a large-scale pretraining corpus. The three new classes (Water, Sky, Roof) are scene-dependent and may be sparse or absent in some frames.

UAVid++ is derived from and builds upon the UAVid dataset. The underlying imagery is from UAVid and its use remains subject to the original UAVid dataset terms and license. If possible, please obtain the source imagery (RGB original frames) from the official release and comply with its terms in addition to the license below.

Licensing

UAVid++ is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

You are free to share and adapt the material for non-commercial purposes, with attribution, and under the same license. Full terms: https://creativecommons.org/licenses/by-nc-sa/4.0/

Use of the underlying UAVid imagery additionally requires compliance with the original UAVid license.

Citation

@article{chiciudean2026uavidplusplus,
  author={Chiciudean, Vivian and Nedevschi, Sergiu and Oniga, Florin},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={UAVid++: Higher-Quality Labels and Expanded Semantic Taxonomy for Aerial Semantic Segmentation}, 
  year={2026},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TGRS.2026.3715191}}

Acknowledgments

This work was supported by the "Romanian Hub for Artificial Intelligence – HRIA" project, Smart Growth, Digitization and Financial Instruments Program, MySMIS no. 351416, Ministry of Investments and European Projects, Romanian Government.

Downloads last month: 71

Total file size:

37.3 GB