Outliers—Do Image and Feature Domain Outliers Coincide in Robotic Applications?

In this work, robotic application data is taken and quantitatively analyzed if outliers in the image domain correlate with outliers in the feature space of CNNs used for the detection of objects. This is necessary to ensure that commonly given reasons for perception failures such as overexposure are valid for any object detection robustness analysis. The outlier detection is done with an encoder-decoder network on images and different layers of different object detection networks. Last, a qualitative analysis of what makes an image an outlier and how this relates to the experience from the application is done.