Closed FedericaGielen closed 7 months ago
Since the goal is to perform the outlier process for occlusion handling, the features for occlusion modeling (loaded here) intend to model the average statistics across the natural images. Therefore, it does not need to be learned, and can simply be computed as the mean feature of randomly sampled natural images.
I am confused about how the parameters are learnt from features in unannotated images in the occlusion modeling.