Supported Datasets

#	Name	H x W	# imgs	Train	Test	Mask	Vessel	OD	Cup	Train-Test split reference
1	Drive	584 x 565	40	20	20	x	x			Staal et al. (2004)
2	STARE	605 x 700	20	10	10		x			Maninis et al. (2016)
3	CHASEDB1	960 x 999	28	8	20		x			Fraz et al. (2012)
4	HRF	2336 x 3504	45	15	30	x	x			Orlando et al. (2016)
5	IOSTAR	1024 x 1024	30	20	10	x	x	x		Meyer et al. (2017)
6	DRIONS-DB	400 x 600	110	60	50			x		Maninis et al. (2016)
7	RIM-ONEr3	1424 x 1072	159	99	60			x	x	Maninis et al. (2016)
8	Drishti-GS1	varying	101	50	51			x	x	Sivaswamy et al. (2014)
9	REFUGE train	2056 x 2124	400	400				x	x	REFUGE
9	REFUGE val	1634 x 1634	400		400			x	x	REFUGE

Add-on: Folder-based Dataset

For quick experimentation we also provide a PyTorch class that works with the following dataset folder structure for images and ground-truth (gt):

root
   |- images
   |- gt

the file names should have the same stem. Currently all image formats that can be read via PIL are supported. Additionally we support hdf5 binary files.

For training a new dataset config needs to be created. You can copy the template ImageFolder and amend accordingly, e.g. the full path of the dataset and if necessary any preprocessing steps such as resizing, cropping, padding etc..

Training can then be started with

bob binseg train M2UNet /path/to/myimagefolderconfig.py -b 4 -d cuda -o /my/output/path -vv

Similary for testing, a test dataset config needs to be created. You can copy the template ImageFolderTest and amend accordingly.

Testing can then be started with

bob binseg test M2UNet /path/to/myimagefoldertestconfig.py -b 2 -d cuda -o /my/output/path -vv