Projects, Carnegie Mellon University, RI, Pittsburgh, USA
For the appearances changes under different weather or season conditions, we proposed a Weather condition invariant unsupervised feature extraction method. By using a conditional cycleGAN method, we achieve to split the weather conditions from the geometry features, without losing the geometry detail at the same time. The extracted weather condition invariant features could make better matching under different weather conditions. The below figure shows the reconstruction result with the extracted geometry features under different conditions. The experiment results could be seen from the below website.