Research map / Label-efficient and robust learning / Domain adaptation and generalization
Adapting Vehicle Detectors for Aerial Imagery to Unseen Domains with Weak Supervision
Xiao Fang, Minhyek Jeon, Zheyang Qin, Stanislav Panev, Celso De Melo, Shuowen Hu, Shayok Chakraborty, Fernando De La Torre
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| stack car, foreground-token and background-token cross-attention maps instead of car map only (ablation) | labeling with YOLOv5 · LINZ->UGRC | AP50: 62.1 -> 68.8 | Improved |
| stack car, foreground-token and background-token cross-attention maps instead of car map only (ablation) | labeling with YOLOv8 · LINZ->UGRC | AP50: 69.8 -> 75.4 | Improved |
| stack multi-channel attention maps instead of car map only (ablation) | labeling with YOLOv5 · DOTA->UGRC | AP50: 68.0 -> 68.5 | No clear effect |
| stack multi-channel attention maps instead of car map only (ablation) | labeling with YOLOv8 · DOTA->UGRC | AP50: 74.5 -> 75.7 | Improved |
| add background token map to car + foreground token maps (ablation) | stacked attention maps (YOLOv5) · LINZ->UGRC | AP50: 63.7 -> 68.8 | Improved |
| add foreground token map to car + background token maps (ablation) | stacked attention maps (YOLOv5) · LINZ->UGRC | AP50: 65.5 -> 68.8 | Improved |
| add background token map to car + foreground token maps (ablation) | stacked attention maps (YOLOv8) · LINZ->UGRC | AP50: 69.1 -> 75.4 | Improved |
| add foreground token map to car + background token maps (ablation) | stacked attention maps (YOLOv8) · LINZ->UGRC | AP50: 73.1 -> 75.4 | Improved |
| fine-tune SD in two stages (tokens fixed in stage two) instead of one stage (ablation) | Stable Diffusion tokens and U-Net · LINZ->UGRC | AP50: 8.5% lower AP50 with one stage | Improved |
| use classifier refinement instead of fixed confidence thresholds 0.3-0.7 (ablation) | pseudo-label filtering (YOLOv8) · UGRC | AP50 | No clear effect |
| use proposed pipeline instead of best unsupervised domain adaptation method | Faster R-CNN based detection · LINZ->UGRC and DOTA->UGRC | AP50: 20.1% and 33.1% | Improved |
| use proposed pipeline instead of best weakly supervised method | Faster R-CNN based detection · LINZ->UGRC and DOTA->UGRC | AP50: 6.2% and 7.2% | Improved |
Closest papers by meaning
- Diagnosing Aerial-View Object Detectors with Foundational Image Generative Models (ECCV 2026 · similarity 0.86)
- DroneFINE: Domain-Aware Parameter-Efficient Fine-Tuning of Vision-Language Detectors for Drone Images (ECCV 2026 · similarity 0.86)
- Progressive Exploration-Conformal Learning for Sparsely Annotated Object Detection in Aerial Images (NeurIPS 2024 · similarity 0.85)
- Multiview Aerial Visual RECognition (MAVREC): Can Multi-view Improve Aerial Visual Perception? (CVPR 2024 · similarity 0.85)
- Scale Optimization Using Evolutionary Reinforcement Learning for Object Detection on Drone Imagery (AAAI 2024 · similarity 0.84)
- DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
