Semantic-aware style transfer for unsupervised aircraft pose estimation
Abstract The aircraft pose estimation based on deep learning has become prevalent in aviation, but its performance heavily relies on expensive and scarce annotated real aircraft data. Although researchers adopt style transfer methods to generate real-style images to reduce dependency on real annotat...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-07-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00171-7 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract The aircraft pose estimation based on deep learning has become prevalent in aviation, but its performance heavily relies on expensive and scarce annotated real aircraft data. Although researchers adopt style transfer methods to generate real-style images to reduce dependency on real annotations, existing methods mainly focus on global style while neglecting local style variations. Crucially, aircraft exhibit distinct local styles that correlate strongly with their semantics. These limitations lead to suboptimal stylization effects, ultimately degrading pose estimation accuracy. To address these challenges, we propose a semantic-aware style transfer method that employs adaptive multi-head attention transfer modules to establish correspondences between local semantics and styles. Meanwhile, we introduce a contextual style loss and a contrastive consistency content loss to guide the network in learning semantic-style relationships, achieving highly realistic stylized effects. Based on this method, we develop an unsupervised aircraft pose estimation framework that relies solely on rendering data and unlabeled real images. The framework employs stylized images to train pose estimator and incorporates a pseudo-supervised segmentation loss to enhance cross-domain feature consistency. This loss utilizes pseudo-labels generated via our unsupervised segmentation approach, which exploits intra-image feature self-similarity and inter-image feature mutual similarity to produce accurate masks. Extensive experiments demonstrate that our style transfer method outperforms existing approaches in both qualitative and quantitative evaluations. Particularly, our unsupervised pose estimation method achieves a 3.39% higher proportion of predictions with errors under 10 $$^\circ $$ ∘ compared to the best supervised baseline. Our method significantly reduces the reliance on manual annotations, thus facilitating the practical deployment of deep learning-based pose estimation in real aircraft scenes. |
|---|---|
| ISSN: | 1319-1578 2213-1248 |