Semi-supervised training of deep neural networks for weather events detection from cameras

Estimating the presence of water on the road is a crucial application when targeting safety-related issues concerning the operational life of a road stretch. Having the capability of deploying a real-time road conditions classifier may represent a powerful way to reduce the number of car accidents caused by bad weather conditions and in general may help making the drivers more aware of the surrounding environment while driving. This work describes the realization of a Machine Learning model able to estimate road conditions from on-site camera images. The project has been carried out within a collaboration between Politecnico di Torino university and Waterview srl. This work describes the construction and the partial annotation of the training dataset starting from the large amount of data collected by the company, but its main focus is the application of self-supervised and semi-supervised techniques for exploiting a large amount of unlabelled data in a supervised task. First, self-supervised training techniques were used to provide a different and possibly better initialization to the model fine tuned to the target task in place of standard Imagenet initialization. In the second stage, the unlabelled portion of the dataset has been labeled by an ensemble of deep neural networks and the provided annotation, as well as the trained deep neural network, have been refined in an iterative approach. Final results show annotations provided to unlabelled data are consistent with those provided by human raters, and performances obtained by the final model are substantially better with respect to the ones obtained by a model trained with labelled frames only.