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

Official code for the IEEE Transactions on Geoscience and Remote Sensing (TGRS) 2026 paper UAVid++. This repository contains the training and inference protocol for the proposed segmentation approach: frozen DINO-pretrained ViT backbones adapted with trainable task-specific heads.

⭐ If you find UAVid++ useful, please consider giving the repo a star and citing our work.

---

Overview

We employ an adaptation strategy that combines frozen DINO-pretrained ViT backbones with trainable task-specific heads for aerial scenarios with limited labeled data. This design preserves the representational power of Vision Transformers while quantifying the gains from improved labels relative to pretrained backbone capacity. Relative to the best-performing state-of-the-art method, our configurations achieve a 2.7% mIoU improvement on UAVid++, up to 12.4% mIoU gains on out-of-distribution UAV datasets, and a 38.85% mIoU gain on a cross-domain dataset.

Method

A frozen DINO-pretrained ViT backbone extracts features that are decoded by a lightweight trainable head. Three head variants are provided:

Head	Description	Params (w/ ViT-H+)
Linear	Single linear projection over patch features	~15.3K
Conv	2× (Conv–BN–ReLU) + 1×1 projection	~22M
UNet-inspired	3 downsample + C/16 bottleneck + 3 upsample with skip connections + 1×1 projection — best	~26M

Repository Structure

After completing the setup, your project directory should look like this:

uavidplusplus-code/
├── code
│   ├── dataloader.py
│   ├── dinov3_wrapper.py
│   ├── inference_strategy.py
│   ├── __init__.py
│   ├── loss.py
│   ├── metrics.py
│   ├── model_inference.py
│   ├── model_train.py
│   ├── __pycache__
│   ├── segmentation_heads.py
│   └── utils.py
├── commands.txt
├── README.md
├── data
│   ├── test
│   ├── train
│   └── val
├── data_processing
│   ├── 1_flatten.py
│   └── 2_preprocess_images.py
├── env.yml
├── models
│   ├── dinov3
│   ├── dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth
│   ├── dinov3_vith16plus_pretrain_lvd1689m-7c1da9a5.pth
│   └── dinov3_vits16_pretrain_lvd1689m-08c60483.pth
└── output 
    ├── 1_9_dinov3_vits_unet_inv_uavid++_mask.png
    ├── 2_9_dinov3_vits_unet_inv_uavid++_colormask.png
    ├── 3_9_dinov3_vits_unet_inv_uavid++_img.png
    ├── 4_9_dinov3_vits_unet_inv_uavid++_overlayed.png
    ├── 5_9_dinov3_vits_unet_inv_uavid++_loss.png
    ├── 6_9_dinov3_vits_unet_inv_uavid++_iou.png
    ├── 7_9_dinov3_vits_unet_inv_uavid++_accuracy.png
    ├── best_model.pth
    ├── test_predictions_9_dinov3_vits_unet_inv_uavid++
    └── training_log_20260718_010918.txt

• output/ – Created automatically when training from scratch. If you are using pretrained models instead, download the weights from 🤗 Hugging Face and place them in this directory.
• models/ – Created after completing Step 1.
• data/ – Created after completing Step 2.

Installation and Running

git clone https://github.com/vivichiciudean/uavidplusplus-code.git
cd uavidplusplus-code

# Python 3.10+ recommended. Make sure CUDA is installed with GPU support.
conda create -n uavidpp python=3.10 -y
conda activate uavidpp
conda install -c conda-forge pyarrow -y
python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
python -m pip install -r req.txt


# Complete Step 1: Download the DINOv3 models
cd ./models
cd ..

# Complete Step 2: Download and prepare the dataset
cd ./data
cd ..

cd code

# Train from scratch
python model_train.py

# Or run inference with pretrained weights
# (see Step 3)
# python model_inference.py

Step 1 – Set up DINOv3

Clone the official DINOv3 repository into the models/ directory:

git clone https://github.com/facebookresearch/dinov3.git

Next, download the DINOv3 checkpoint(s) you want to use from the official Hugging Face collection:

https://huggingface.co/collections/facebook/dinov3

Place the downloaded checkpoint files directly in the models/ directory (alongside the dinov3/ repository).

Step 2 – Download and Preprocess the Dataset

Download the UAVid++ dataset and review the dataset documentation on 🤗 Hugging Face.

This guide assumes you are using the full UAVid++ dataset. If you choose a different dataset variant, you may need to adjust the preprocessing steps (and the dataloader) accordingly.

After downloading the dataset, preprocess it in the following order:

1. Run 1_flatten.py separately for each split:

   • uavid_train/
   • uavid_val/
   • uavid_test/

2. Run 2_preprocess_images.py to generate image crops for the:

   • training split (uavid_train)
   • validation split (uavid_val)

Before running either script, ensure your current working directory is correct. Also review the TODO comments in the preprocessing scripts and update any required paths or configuration before execution.

Step 3 (Optional) – Use Pretrained Weights

If you do not want to train the model from scratch, you can use one of the provided pretrained checkpoints instead.

Pretrained checkpoints are available on 🤗 Hugging Face alongside the dataset. We currently provide checkpoints for the following configurations:

• ViT-S UNet trained on UAVid++
• ViT-H+ UNet trained on UAVid++
• ViT-7B UNet trained on UAVid++

Download the desired checkpoint and place it in the output/ directory. You can then run inference directly using model_inference.py.

If you use a different DINOv3 backbone, you will need to train the model from scratch.

Training and Inference

Before training or running inference, complete the setup described in Step 1 and Step 2. If you plan to use pretrained segmentation models, also complete Step 3.

Training Settings

The experiments reported in the paper use the following configuration:

Setting	Value
Optimizer	AdamW
Base learning rate	3 × 10⁻⁵
Weight decay	0.001
Batch size	8
Epochs	40
LR schedule	Linear warmup (5 epochs), then cosine annealing
Augmentations	Geometric + photometric + coarse dropout
Input tiling	Non-overlapping 1088 × 1088 tiles; 8 tiles per 4K frame (1,600 training + 560 validation tiles)
Hardware	1 × NVIDIA A100 40 GB

Training

After selecting the DINOv3 backbone (e.g., ViT-H+) and the corresponding experiment configuration. The training script also performs inference.

python model_train.py

Inference

To run only inference using a trained or pretrained checkpoint:

python model_inference.py

Inference uses the same 1088 × 1088 non-overlapping tiling strategy as training. Predictions from individual tiles are stitched together into a full-resolution segmentation map. Performance is reported using per-class IoU and mean IoU (mIoU).

Model Zoo & Main Results

We provide pretrained checkpoints for the following configurations:

• DINOv3 ViT-7B + UNet (best-performing model)
• DINOv3 ViT-H+ + UNet
• DINOv3 ViT-S + UNet (smoke test)

Proposed approach on UAVid++ (mIoU %, train/test on UAVid++):

Backbone	Head	mIoU
DINOv3 ViT-L	UNet	80.82
DINOv3 ViT-H+	Linear	70.13
DINOv3 ViT-H+	Conv	80.86
DINOv3 ViT-H+	UNet	81.42
DINOv3 ViT-7B	UNet	82.04

Head ablation (mIoU %, Linear / Conv / UNet):

Backbone	Linear	Conv	UNet
DINOv3 ViT-L	72.42	80.41	80.82
DINOv3 ViT-H+	70.13	80.86	81.42
DINOv3 ViT-7B	74.95	81.99	82.04

Generalization

Out-of-distribution UAV datasets (trained on UAVid++, tested on UDD and VDD - using IDD labels), mIoU %:

Test set	D2LS	Ours (ViT-H+ UNet)
UDD	55.46	67.89 (+12.4)
VDD	60.14	69.00 (+8.9)

Cross-domain (trained on UAVid++, tested on OpenEarthMap), mIoU %:

Method	mIoU
D2LS	24.10
Ours (ViT-H+ UNet)	62.95 (+38.85)

License

This code and the UAVid++ annotations are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0) — https://creativecommons.org/licenses/by-nc-sa/4.0/.

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={64},
  number={},
  pages={5634816-5634816},
  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.