SIMPL Multi-Aspect MADD: Rapidly Generating Low-Cost Multi-Aspect Military Data for All-Domains at Scale

In this work we extend an approach known as SIMPL (Synthetic object IMPLantation) to construct a large and diverse synthetic dataset known as MADD (Military All-Domain Dataset). Our extension to SIMPL provides the ability to easily and rapidly generate a massive data set to overcome issues in object recognition algorithm development like: limited data, limited object viewing aspects, data collection infeasibility, or zero-shot problems. In total MADD contains 100 unique targets positioned at a total of 100,000 times in diverse backgrounds. Each target is captured from >50 unique viewing aspects, providing exceptional coverage of what an observing algorithm may see in the wild. We showcase the effectiveness of our method for building or augmenting deep neural networks where there is no real-world data available (Zero-Shot scenarios), and we also demonstrate the dataset's sensitivities using our novel MADD-500 benchmark test set.

Duke Scholars

Author Leslie M. Collins Electrical and Computer Engineering