Legacy code recovered from the project done for the webly-supervised network paper.
Please see the project website here, which includes models and data.
The code is divided into two parts:
layersfolder contains the loss function implemented to encode the relationship graph. Specifically, it computes the cross entropy between the soft-max output and the smoothed target category.
subdiscoverfolder contains the code used to extract
fc7features for subcategory discovery -- finding bounding boxes of a category given weak, noisy (that is, webly) supervision of category labels for the entire image.