search

WenqiangX / ese_seg

Explicit Shape Encoding for Real-time Instance Segmentation
105 stars 16 forks source link

Mask RCNN #1

Closed yyfyan closed 5 years ago

yyfyan commented 5 years ago

hello, best job. I noted Mask RCNN achieved 68.5 mAP on the VOC12 val set, how train the Mask RCNN on the VOC dataset? the train set including SBD+VOC12?

WenqiangX commented 5 years ago

Hi, VOC 2012 val set is used for validation, training including sbd + voc12. It is a common practice since SDS.

yyfyan commented 5 years ago

Thanks. the training set size is 10582? @WenqiangX

WenqiangX commented 5 years ago

@yyfyan no the training set from sbd is http://home.bharathh.info/pubs/codes/SBD/train_noval.txt, this will exclude all the val in voc2012. The default train/val split will contain ~500 voc2012 val images in train split.