Relation networks for few-shot video object detection

Abstract

This paper describes a new few-shot video object detection framework that leverages spatio-temporal information through a relation module with attention mechanisms to mine relationships among proposals in different frames. The output of the relation module feeds a spatio-temporal double head with a category-agnostic confidence predictor to decrease overfitting in order to address the issue of reduced training sets inherent to few-shot solutions. The predicted score is the input to a long-term object linking approach that provides object tubes across the whole video, which ensures spatio-temporal consistency. Our proposal establishes a new state-of-the-art in the FSVOD500 dataset.