Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision

How can we use synthetic data to improve cross-domain object detection performance in aerial imagery?

We harness the power of Stable Diffusion to generate synthetic images and automatically annotate them using cross-attention maps.

Our method outperforms supervised learning on source domain data, weakly supervised adaptation methods, unsupervised domain adaptation methods, and open-set object detectors by 4-23%, 6-10%, 7-40%, and more than 50% AP₅₀, respectively.

We introduce two large-scale datasets captured in Utah and Selwyn to advance vehicle detection in aerial imagery.

Method Overview

We propose a pipeline that generates high-quality aerial images along with their labels to address the domain shift challenge in aerial image analysis. The first stage involves finetuning Stable Diffusion to synthesize images for both the source and target domains. The second stage then automatically generates labels for the synthetic target images by leveraging cross-attention maps. The effectiveness of this pipeline is validated by a significant boost in detector performance. This approach surpasses both baseline detectors trained on source images and open-set detectors inferred directly on the target domain. Please see our paper for more implementation details.

Experiment Results

Cross-domain object detection Comparison with VLLMs Cross-attention maps

Method	Backbone	AP50 (%)
		LINZ→UGRC	DOTA→UGRC
Supervised Object detection (Source-only)
Faster-RCNN	Faster R-CNN	53.1	52.5
YOLOv5	YOLOv5	63.2	57.6
YOLOv8	YOLOv8	62.9	71.4
ViTDet	ViTDet	55.7	50.4
Open-set object detection
GLIP-T	Swin-T	8.7	8.7
OmDet-Turbo	Swin-T	14.4	14.4
OWL-V2	CLIP ViT	17.9	17.9
Unsupervised domain adaptation object detection
SIGMA	Faster R-CNN	36.2	34.8
TIA	Faster R-CNN	39.0	42.4
Adapt. Teacher	Faster R-CNN	29.0	37.0
CycleGAN-Turbo	YOLOv5	29.1	60.8
SSDA-YOLO	YOLOv5	52.3	49.6
Cross domain weakly supervised object detection
OCUD	Faster R-CNN	63.1	65.3
H2FA R-CNN	Faster R-CNN	61.8	68.3
Ours	Faster R-CNN	69.3	75.5
Ours	YOLOv5	68.8	68.5
Ours	YOLOv8	75.4	67.1
Ours	ViTDet	72.0	67.1

Cross-domain object detection results. LINZ to UGRC and DOTA to UGRC. We report the AP₅₀ result.

Method	Vision Backbone	LINZ→UGRC Precision(%)	LINZ→UGRC Recall(%)
Vision Large Language Model
Gemini 1.5 Flash	-	2.9	44.5
Gemini 2.0 Flash-Lite	-	6.6	26.3
InternVL3-8B	InternViT	4.7	22.0
Qwen2.5-VL-7B	ViT	0.4	4.8
DeepSeek-VL2-Tiny	SigLIP-SO400M	9.2	26.8
LLaVA-NeXT	CLIP ViT	5.5	4.7
Ours
Ours	Faster R-CNN	63.8	68.2
Ours	YOLOv5	67.2	67.3
Ours	YOLOv8	70.0	76.3
Ours	ViTDet	72.0	67.1

Table 3. Comparison between our methods and VLLMs on UGRC dataset. We report the precision and recall metrics.

Cross-attention maps for different tokens. We analyze the effectiveness of utilizing multi-channel cross-attention maps by assessing the label quality on synthetic UGRC images. (a) Synthetic UGRC images. (b) Labels generated using only the cross-attention map of the word “car”. (c) Labels generated by using multi-channel heatmaps. (d) Multi-channel cross-attention maps. (e) Cross-attention maps of the word “car”. (f) Cross-attention maps of token [V₁], designed to capture the concept of cars. (g) Cross-attention maps of token [V₃] for background, inverted for comparison. Heatmaps highlight attention intensity. In (b) and (c), dotted lines denote predicted pseudo-boxes; white dots = vehicle centers.

LINZ and UGRC Dataset

We introduce two large-scale and high-quality benchmarks, LINZ and UGRC, designed to evaluate domain adaptive object detection (DAOD) methods under realistic distribution shifts. These datasets support geographic domain shift evaluation, and provide dense, expert-level annotations. Together, these datasets comprise over 5 million annotated images and over 80,000 labeled object instances, establishing a robust foundation for testing DAOD algorithms.

BibTeX

@misc{fang2025adaptingvehicledetectorsaerial,
      title={Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision}, 
      author={Xiao Fang and Minhyek Jeon and Zheyang Qin and Stanislav Panev and Celso de Melo and Shuowen Hu and Shayok Chakraborty and Fernando De la Torre},
      year={2025},
      eprint={2507.20976},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.20976}, 
}