REVIVAL: Retrieval-Enhanced Vetting and Identification for Visual Active Learning

Computer vision detector algorithms rely on large volumes of labeled data to learn robust, generalizable representations for objects and features present within satellite imagery. Machine-assisted labeled techniques have made it possible to annotate massive datasets by leveraging pre-trained models to generate labels at scale. However, automated labeling tends to amplify the presence of common objects of interests (OOIs) while underrepresenting scare, rare, challenging, or low-frequency OOIs. Capitalizing on these objects at the tail of the distribution is critical for developing reliable geospatial AI/ML systems that generalize across diverse environments and global contexts. To address this problem, we proposeáREVIVAL: Retrieval-Enhanced Vetting and Identification for Visual Active Learning. This machine-assisted labeling platform aims to improve the accuracy and efficiency of annotating rare or challenging OOIs in overhead satellite imagery. REVIVAL uses advanced uncertainty quantification techniques to flag regions where the models exhibit ambiguity or low confidence in its detections. Additionally, REVIVAL implements a scalable similarity search engine to retrieve contextually similar imagery to find supplementary examples of the OOIs. Lastly, REVIVAL includes an intelligent model-assisted labeling interface that prioritizes high-value human input for the confirmation and refinement of the labels produced from our framework. This integrated system will enable the rapid discovery, clustering, and annotation of challenging OOIs to accelerate the creation of balanced and representative training datasets for computer vision algorithms in the remote sensing domain.