Abstract: The future of sophisticated robotic automation will rely upon several specialized machine learning models, sometimes referred to as a mixture of experts or multi-stage models, that work in tandem to achieve a goal. The value of these complex systems comes from improved performance in meeting the objective. We will detail an example of how multi-stage models can be used for accurate predictions on high-resolution aerial imagery ingested and processed in real-time. The input in this case is a high-resolution video feed from an unmanned aerial vehicle (i.e. a drone). Achieving accurate predictions on high-resolution imagery is difficult when the input is resized, which is a common practice for many well-known computer vision models. We’ll use two models that specialize in localization and classification, respectively. We will observe how this divide-conquer strategy enables accurate classifications of high-resolution images and reduces localization error.
Bio: Neela is a Data Scientist in the Analytics Team at SAS. She primarily focuses on developing tangible, end-to-end AI applications using deep learning with an emphasis on Computer Vision and creating machine learning workflows to integrate open source within SAS analytics. She also works on designing content to articulate this process for customers in solving their analytical needs.