Adaptive image zoom-in with bounding box transformation for UAV object detection

2026-02-07

Isprs Journal of Photogrammetry and Remote Sensing

AI Summary

This work addresses the challenge of detecting small and sparsely distributed objects in drone-captured images, which often limits the performance of general-purpose detectors. To this end, the authors propose an adaptive non-uniform scaling framework that learns to locally magnify foreground regions through a lightweight offset prediction mechanism, coupled with a corner-aligned spatial transformation of bounding boxes. This approach enables end-to-end, architecture-agnostic feature enhancement and detection. By introducing a bounding-box-based scaling objective function, the method achieves significant performance gains on the VisDrone, UAVDT, and SeaDronesSee benchmarks, with an improvement of over 8.4 mAP on SeaDronesSee while adding only approximately 3 ms to inference latency.

Technology Category

• Computer Vision: Object Detection & Categorization
• Search and Optimization: Learning to Search
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Problem

• UAV object detection
• small object detection
• image zoom-in
• bounding box transformation
• adaptive zooming

Innovation

• adaptive zoom-in
• bounding box transformation
• UAV object detection
• non-uniform image warping
• small object detection