Comparison of ID performance using real and synthetic imagery
Abstract
Recent experiments performed at the US Army Night Vision and Electronic Sensors Directorate (NVESD) provide significant insight into the validation of synthetic imagery for use in human perception experiments. This paper documents the procedures and results of target identification (ID) experiments using real and synthetic thermal imagery. Real imagery representing notional first generation and advanced scanning sensor systems was obtained. Parameters derived from the sensor data were used to generate synthetic imagery using the NVESD Paint the Night simulation. Both image sets were then used in a target identification experiment with trained human observers. Perception test results were analyzed and compared with metrics derived from the imagery. Several parameters missing from the original truth data were found to correlate with differences in the perception data. Synthetic data were regenerated using these additional parameters. A subsequent perception experiment confirmed the importance of these parameters, and a good match was obtained between real and synthetic imagery. While the techniques used in this series of experiments do not constitute a definitive method for validating synthetic imagery, they point to some important observations on validation. The main observation is that both target and local background characteristics must be sufficiently specified in the truth data in order to obtain good agreement between synthetic and real data. The paper concludes with suggestions as to the level of detail necessary for truth data when using synthetic imagery in perception experiments.
Publication Title
Proceedings of SPIE - The International Society for Optical Engineering
Recommended Citation
Jacobs, E., Edwards, T., Miller, B., & Hodgkin, V. (2002). Comparison of ID performance using real and synthetic imagery. Proceedings of SPIE - The International Society for Optical Engineering, 34-41. https://doi.org/10.1117/12.477465
