Streamlined hybrid annotation framework using scalable codestream for bandwidth-restricted UAV object detection

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

The success of emergency missions depends on the rescuers' ability to intervene quickly. A work increasingly facilitated by unmanned aerial vehicles, which provide crucial visual information on the situation. However, the communication capacity is often constrained by a limited bandwidth, preventing fast transmission and thereby delaying the quick decision-making necessary in emergency situations. To address these challenges, this master thesis presents a hybrid annotation framework adapted for emergency situations that takes advantage of the JPEG 2000 compression algorithm. The framework is designed to facilitate object detection tasks with constrained bandwidth. Fast annotation is ensured by a fine-tuned deep learning model working at low-resolution, and reliability is provided by the correction of human annotator experts in enhanced resolution areas of uncertainty. It has been shown that the proposed framework decreases the response time up to 15 times while maintaining satisfactory performance, losing only 10% compared to a hybrid baseline approach.