PASCAL VOC 2012 dataset

[sunggukcha]

Your paper and 'splits.ipynb' describe PASCAL VOC 2012 train set has 11685 images, while your currently distributing code includes 10582 images for training (check your data/datasets/voc12/split/train_list.npy). I think 10582 is the usual augmented PASCAL VOC 2012 training set size. Where the thousand of training images come from?

Please also check 'data/datsets/voc12/split/seenval_cls.npy' which is missing.