Few-Shot Object Detection on Remote Sensing Images
The aim of this project is to achieve the best accuracy possible on novel object with very limited samples in the 1-5-10 shots scenario. As remote sensing images are used, object refers to all kind of man-made element, from train station to vehicles. To improve the results, a sub-part approach has been investigated by having a contrastive training procedure. This work lead to an Conference paper that has been presented at IGARSS 2023. Further details are available on the associated GitHub page.