Investigating the Sim-to-Real Generalizability of Deep Learning Object Detection Models

Abstract

State-of-the-art object detection models need large and diverse datasets for training. As these are hard to acquire for many practical applications, training images from simulation environments gain more and more attention. A problem arises as deep learning models trained on simulation images usually have problems generalizing to real-world images shown by a sharp performance drop. Definite reasons and influences for this performance drop are not yet found. While previous work mostly investigated the influence of the data as well as the use of domain adaptation, this work provides a novel perspective by investigating the influence of the object detection model itself. Against this background, first, a corresponding measure called sim-to-real generalizability is defined, comprising the capability of an object detection model to generalize from simulation training images to real-world evaluation images. Second, 12 different deep learning-based object detection models are trained and their sim-to-real generalizability is evaluated. The models are trained with a variation of hyperparameters resulting in a total of 144 trained and evaluated versions. The results show a clear influence of the feature extractor and offer further insights and correlations. They open up future research on investigating influences on the sim-to-real generalizability of deep learning-based object detection models as well as on developing feature extractors that have better sim-to-real generalizability capabilities.