.png){kind=icon}
.png){kind=icon}
Return
Towards Unified Transformer for UAV-Based Multi-Task Oriented Object Detection
.png){kind=icon}
OA
AI
DOI：10.1049/ipr2.70343.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
UAV-based visible-thermal object detection is crucial for all-weather aerial perception. However, prevailing methods that train separate models for visible, thermal, and fusion tasks incur significant computational redundancy. Moreover, under the typical UAV perspective, objects often appear with arbitrary orientations, making oriented bounding boxes more suitable than horizontal ones for accurate localization. To address both the efficiency challenge and the need for precise orientation-aware detection, we propose TUMODet, a unified multi-task oriented detection framework. Built upon the spatial transform decoupling (STD) detector, TUMODet incorporates task-token prompt embedding and static weighted fusion, enabling a single model to dynamically handle three detection tasks (visible, thermal, and fused modalities). Comprehensive experiments on the DroneVehicle dataset demonstrate that our unified framework significantly improves parameter efficiency and resource utilization while achieving competitive performance across all tasks, particularly in the challenging fusion scenario. The results validate TUMODet as an efficient and practical solution for unified multimodal detection on resource-constrained UAV platforms.
Keywords:
multi-task learning
oriented object detection
task-token prompt
transformer
UAV
visible-thermal object detection
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
2.2
Papers:
118
Citations:
4.7K