Presenting results for a simple, unsupervised road segmentation algorithm. The underlying data is from GTA 5. A one layer autoencoder network was learnt and was used to transform each frame of the top left video into the color model in the top right video. A number of these models were combined to form the bottom right video. A set of handcrafted filter sets are used to obtain the bottom left video. The yellow colored region is hypothesized to be ‘safe’ for driving while random color pixels in their midst is the ‘anomaly’ or ‘pothole’. Note that the entire algorithm is completely unsupervised and there is no template provided to the algorithm.
Home » Unsupervised Road Segmentation