dataset composition about pfenet

[tianzhuotao]

We follow the instructions in http://home.bharathh.info/pubs/codes/SBD/download.html to use sbd_train that excludes all images in the VOC validation set. The sbd_val set overlaps with the validation set of VOC.

The final list (5900) is derived from the SBD train (5623) + VOC train (1464) by removing all duplicate images.

[roywithfiringblade]

Thanks for your reply, i just check the setting of CANet, after removing the small object and duplicate images, the training set contains 10092 images, which is far more than your settings.
I think it is the subset of 'trainaug.txt' in deeplab which contains 10582 no duplicated images( voc_train(1464)+sbd_train(6820)+sbd_val(2298)). Intuitively, a bigger train_set may boost the performance. I will try to train your model in that way and check the performance.Thanks!

[tianzhuotao]

Thank you.

Strictly speaking, because few-shot segmentation uses the validation set of VOC for evaluation, images contained in the validation set should be excluded. It is also what SBD webpage emphasizes. But, I guess the performance gap will not be very large because few-shot segmentation evaluates only the novel classes. If so, training with 10582 images (trainaug) is also okay.