École Nationale Supérieure de
Techniques Avancées

Abstract

This report investigates few-shot object detection for high-resolution remote sensing imagery, addressing the challenge of detecting objects classes with limited annotated data. Building on a prototype-based method [1], we conduct a reproduction study to assess its performance, analyse its strengths and limitations, and evaluate its adaptability across different datasets. We show that a more rigorous evaluation of the method produces slightly worse results, in particular when adding a prototype fine-tuning step. This thereby confirms the sensitivity of frugal learning to the chosen example sets. We evaluate the method’s components in isolation and on more datasets, and use this evaluation to drive two extensions of the method: improving the feature maps resolution with a joint bilateral upsampling scheme, and then producing detections directly from the feature maps without a region proposal network