Zero-Shot Detection in Satellite Images

The research is devoted to the adaptation of the Zero-Shot Detection paradigm for the task of detecting objects in satellite images. The aim of the work was to experimentally test the hypothesis that it is possible to detect classes missing in the training sample based on semantic comparison of visual features and textual descriptions. The methodology involved fine-tuning the detectors (Faster R-CNN, YOLOv10x) on a dataset of urbanized objects (4 classes, 1300 images), generating textual descriptors for “new” classes (“fire”, “smoke”) using GPT-4o and creating a shared semantic space via ResNet-50 (visual) and GPT-2 (textual) embeddings. The results confirmed the performance of zero-shot detection: the models achieved APs of 0.521-0.548 (fire) and 0.493-0.512 (smoke) without training on the corresponding data.