Low-Shot Detection in Remote Sensing Imagery

With the ever-growing number of imaging satellites in orbit the job of an analyst will change from eyes on pixels to analysis of informationfrom imagery thanks to automated image processing like object detection and change detection. Object detection algorithms have advancedto near human performance given that there is sufficient labeled data on which to train; however, obtaining this data is costly and timeconsuming. On the other hand the mission is immediate and potentially urgent. A preferable solution is to build upon historical training databut enable object detectors to identify new high value objects of interest from small sets of new data labels, i.e. new low shot classes. Etegentproposes to extend their demonstrated dual network low shot solution as well as investigate cutting edge approaches that leverage contextto provide a comprehensive solution to the low shot problem.