search

rowanz / neural-motifs

Code for Neural Motifs: Scene Graph Parsing with Global Context (CVPR 2018)
https://rowanzellers.com/neuralmotifs
MIT License
525 stars 115 forks source link

About the usage of flags #15

Closed hellochick closed 6 years ago

hellochick commented 6 years ago

Hey! Really appreciate this amazing work. I have a little question about these four flags: -gt_box, -old_feat, -pass_in_obj_feats_to_decoder, -pass_in_obj_feats_to_edge. Could you please explain these flags? Thank you a lot.

Btw, I am trying to use MaskRCNN to extract features and predict the relations between instances. Is there any suggestions? Hope to hear from you soon!

Best wishes, HsuanKung

rowanz commented 6 years ago

Sure! the only one that really matters is -gt_box as that controls whether we train/eval on ground truth boxes or on predicted boxes. -old_feats doesn't do anything right now. -pass_in_obj_feats_to_decoder/edge controls whether we want to duplicate the object features (from Faster RCNN) for object/edge classifications. Anyways, you might want to look at the scripts directory, which has the flags I used.

MaskRCNN is built on top of Faster RCNN, but it doesn't make sense to train it on this dataset as Visual Genome doesn't have segmentations. It also just does object detection/segmentation, so not relations. You'll need another model on top of the detector for that (like MotifNet or one of the baselines). hope that helps!