Self-Aligned Spatial Feature Extraction Network for UAV Vehicle Reidentification

Compared with existing vehicle reidentification (VeID) tasks conducted datasets collected by fixed surveillance cameras, VeID for an unmanned aerial (UAV) is still under-explored and could be more challenging. Vehicles the same color type show extremely similar appearances from UAV’s perspective so that mining fine-grained characteristics becomes necessary. Recent works tend to extract distinguishing information regional features component features. The former requires input images aligned latter entails detailed annotations, both of which are difficult meet in UAV application. To efficient avoid tedious annotating work, this letter develops unsupervised self-aligned network consisting three branches. introduced a self-alignment module convert variable orientations uniform orientation, implemented under constraint triple loss function designed spatial On basis, features, obtained vertical horizontal segmentation methods, global integrated improve representation ability embedded space. Extensive experiments on UAV-VeID dataset, our method achieves best performance compared recent (ReID) works.