Impact Factor:12.2
DOI number:10.1016/j.isprsjprs.2025.07.024
Journal:ISPRS Journal of Photogrammetry and Remote Sensing
Key Words:Thermal Infrared Remote Sensing Multiple Angles Object Detection
Abstract:Multi-angle thermal infrared (MATIR) remote sensing provides valuable day-night and multiple angular information that is of significant practical value in applications. However, multi-angle data heterogeneity is one of the core challenges in feature learning and scene understanding, which could severely degrade the model inference performance of deep neural networks. To address this issue, this study proposes a new fine-grained dataset and a unified method for the MATIR object detection task. In detail, the fine-grained MATIR object detection (MATIR-OD) dataset is captured by an unmanned aerial vehicle (UAV)-based platform, which offers significant advantages in terms of cost efficiency and exceptional maneuverability. The MATIR-OD dataset comprises 24 fine-grained and multi-angle data subsets, containing a total of 43,540 instances. Moreover, the unified MATIR object detection method, denoted as U-MATIR, includes the heterogeneous label space module and hybrid view cascade module. In the multi-angle object detection task, based on four public datasets and the proposed dataset, the all-angle experimental results show that the U-MATIR outperforms the ground- or aerial-view object detection models, increasing accuracy with an approximately 18–65% improvement in the mean Averaged Precision (mAP) metric, which exhibits notable robustness and generalization ability. In addition, the extensive experiments demonstrate the boundaries of robustness and generalization ability under 20–120 m and 30–90° fine-grained observation data. In particular, the optimal detection angle is defined as 60° under the above observation heights. The MATIR object detection dataset and unified method provide new insight for accurate multi-angle localization and achieve competitive detection performance.
Indexed by:Journal paper
Discipline:Engineering
First-Level Discipline:Surveying and Mapping
Document Type:J
Volume:228
Page Number:438-452
Translation or Not:no
Included Journals:SCI、EI
Links to published journals:https://www.sciencedirect.com/science/article/pii/S0924271625002862
First Author:Chenchen Jiang
Correspondence Author:Huazhong Ren
All the Authors:Fengguang Li,Zhonghua Hong,Hongtao Huo,Junqiang Zhang,Jiuyuan Xin
Date of Publication:2025-07-17