SBIR-STTR Award

Etegent proposes Automated Learning from Unsupervised Repositories of Data (ALURD). ALURD incorporates a trained detector to feed a semi-supervised discrimination apparatus that leverages state-of-the-art approaches.in semi-supervised learning (SSL).  The need for automated labelling of overhead data is obvious, less obvious is that these unlabelled images provide an opportunity to improve autonomous labelers making them more accurate and more dynamic.  Extracting even a small amount of information from the stream of unlabelled samples has the potential to massively impact the quality of machine learners for remotely sensed imagery. The proposers intend to improve classification in satellite imagery with limited annotations.   \n\n Etegent proposes Automated Learning from Unsupervised Repositories of Data (ALURD). ALURD incorporates a trained detector to feed a semi-supervised discrimination apparatus that leverages state-of-the-art approaches.in semi-supervised learning (SSL).  The need for automated labelling of overhead data is obvious, less obvious is that these unlabelled images provide an opportunity to improve autonomous labelers making them more accurate and more dynamic.  Extracting even a small amount of information from the stream of unlabelled samples has the potential to massively impact the quality of machine learners for remotely sensed imagery.  The proposers intend to improve classification in satellite imagery with limited annotations.

The need for automated labelling of overhead data is obvious, less obvious is that these unlabelled images provide an opportunity to improve autonomous labelers making them more accurate and more dynamic. Extracting even a small amount of information from the stream of unlabelled samples has the potential to massively impact the quality of machine learners for remotely sensed imagery. The proposers intend to improve detection and classification in satellite imagery with limited annotations. In addition to improved classification for currently identified targets of interest, the proposers intend to develop methods to automate the process of novel class discovery. Instead of developing a "closed world" system, ALURD will be constructed to identify targets not contained in the training set and flag them for manual annotation. Additionally, imagery containing targets which the system has difficulty classifying will be flagged for human interrogation.