Quantifiable Robustness Estimation for Object Detection with CNNs Using Intrinsic Dimensionality

Interpreting the results of Convolutional Neural Networks remains a challenging task. Quantitative evaluations apart from precision, recall, and their extensions are rare and usually do not cover the necessary aspects of specific applications. In this work, a methodology based on the intrinsic dimensionality of the image space and latent space in multiple layers is presented. This methodology has been used in other literature for classification but is leveraged to object detection where the interpretation of the results is more complex. The suitability of the intrinsic dimensionality is evaluated first for general augmentation techniques in multiple datasets and with multiple networks and later on a specific use case with multiple disturbances included. With the help of the intrinsic dimensionality, conclusions about the robustness can be drawn which are not apparent from the precision and the suitability of the methodology as an auxiliary quantifiable metric therefore shown.