Robustness and Deployability of Deep Object Detectors in Autonomous Driving

Abstract

Many successful deep learning networks for object detection have been proposed in recent years, but their direct comparison is difficult due to different resolutions, training and validation data and frameworks used to train these networks. Moreover, suitability of these object detectors in context of autonomous driving is not compared or studied extensively. Autonomous driving has a wide operational domain containing diverse open world scenarios like different weather, daylight and road conditions etc. Most of the object detection datasets are from specific operational domains, e.g. KITTI dataset is mostly urban scenarios. In this work, we aim to first club many open source datasets (simulated and real) available for object detection in autonomous driving scenario. Clubbing the datasets makes them diverse by including different weather, lighting, time of the day and road conditions. Secondly, by clubbing most of the publicly available object detection datasets, we show the generalization ability of object detection networks in diverse scenarios. We also provide comparative study of accuracy vs resolution and the results of DNN based object detection on real images by training only on simulated data. Additionally, range of detection is really important for autonomous driving basic functionality like highway driving, and market trend is to move towards higher resolution images, yet most of the well known DNN based object detectors operate at very small image resolution which restricts range of detection. We demonstrate that the use of ROI instead of entire image can help in increasing detection range. Finally, in the context of autonomous driving these DNNs needs to run on low powered Edge devices. We measure the deployability of these DNNs by measuring their run time performance on Nvidia Drive PX2.