Benchmarking performance of object detection under images distortions of an uncontrolled environment

International audience; The robustness of object detection algorithms plays a prominent role in real-world applications, especially in the uncontrolled environments due to distortions during image acquisition. It has been proved that object detection methods suffer from in-capture distortions to perform a reliable detection. In this study, we present a performance evaluation framework of the state-of-the-art object detection methods on a dedicated dataset containing images with various distortions at different levels of severity. Furthermore, we propose an original strategy of image distortion generation applied to the MSCOCO dataset that combines some local and global distortions to reach a better realism. We have shown that training with this distorted dataset improves the robustness of models by 31.5%. Finally, we provided a custom dataset including the natural images distorted from MS-COCO to perform a more relevant evaluation of the robustness concerning distortions. The database and the generation source codes of the different distortions are publicly available.