Mask_rcnn: Difference of skimage.transform.resize and scipy.misc.imresize
I noticed latest code change to use skimage.transform.resize, but seems for model trained using either old code(with scipy.misc.imresize) or latest code.
I always get better nuclei detection map score if change to use scipy.misc.imresize for detection.
ie, using latest training, with mAP 0.5800 but if I change to use scipy.misc.imresize for detection it can improve to 0.58288.
chenghuige
>All comments
I think we probably should use scipy.misc.imresize. Based on https://github.com/python-pillow/Pillow/issues/2865, and http://scikit-image.org/docs/dev/api/skimage.transform.html#skimage.transform.resize, it seems that skimage.transform.resize is not using anti-aliasing by default, which will create artifact when down sampling. It seems scipy.misc.imresize has that implemented. I think 0.14 version of skimage.transform.resize will have anti-aliasing implemented, but we need to install from their source. scipy.misc.imresize will deprecate soon, so skimage.transform.resize will be the future and hope it catch up in time.
RLstat
on 11 Apr 2018
Related issues
PaulChongPeng
路
4Comments
Mabinogiysk
路
3Comments
canerozer
路
3Comments
taewookim
路
4Comments
chrispolo
路
4Comments
Most helpful comment
I think we probably should use scipy.misc.imresize. Based on https://github.com/python-pillow/Pillow/issues/2865, and http://scikit-image.org/docs/dev/api/skimage.transform.html#skimage.transform.resize, it seems that skimage.transform.resize is not using anti-aliasing by default, which will create artifact when down sampling. It seems scipy.misc.imresize has that implemented. I think 0.14 version of skimage.transform.resize will have anti-aliasing implemented, but we need to install from their source. scipy.misc.imresize will deprecate soon, so skimage.transform.resize will be the future and hope it catch up in time.