Solomon, J. and Charette, F., "A Bootstrap Approach to Training DNNs for the Automotive Theater," SAE Technical Paper 2017-01-0099, 2017, doi:10.4271/2017-01-0099.
The proposed technique is a tailored deep neural network (DNN) training approach which uses an iterative process to support the learning of DNNs by targeting their specific misclassification and missed detections. The process begins with a DNN that is trained on freely available annotated image data, which we will refer to as the Base model, where a subset of the categories for the classifier are related to the automotive theater. A small set of video capture files taken from drives with test vehicles are selected, (based on the diversity of scenes, frequency of vehicles, incidental lighting, etc.), and the Base model is used to detect/classify images within the video files. A software application developed specifically for this work then allows for the capture of frames from the video set where the DNN has made misclassifications. The corresponding annotation files for these images are subsequently corrected to eliminate mislabels. The corrected annotations and corresponding images are then collated and used to re-train the base DNN model. The process is subsequently repeated, where the newly trained model, (which we will refer to as the Cycle 1 model), is used to review the same subset of video files. The process is repeatedly indefinitely until a satisfactory level of accuracy is achieved, where each cycle of training results in an incrementally improved model. The ultimate objective is to create a robust DNN able of correctly detecting and classifying a large range of objects pertinent to the general driving experience.