부산대학교 도서관

This contribution discusses the use of variational data synthesis as a tool to analyze and understand limitations of performance of DNNs (deep neural networks) in perception tasks. To date, no universally accepted methodologies for validating ML (Machine Learning) -based perception exist. Instead of aiming for the randomized acquisition of huge amounts of validation data, either from real world capture or from simulation, we propose a guided concept to analyze perception performance using systematic parameter variations. The concept is based on parameterized, generative content used for data synthesis in our validation engine. The latter is composed of the actual data synthesis module, automated execution and evaluation of the perception function under test and a control module, which allows specification of parameter variation towards a validation goal. Further we investigate the use of physical parameters, like object occlusion rates and pixel area for the identification of critical cases for perception. We present experiments for semantic segmentation of pedestrians in an urban environment using two different DNN algorithms.