I am trying to run your graph-rcnn on a custom scene-graph dataset. I had some questions regarding how __getitem__ in dataloader creates the targets. I noticed that you create a 2D array obj_relations with the shape of number of bboxes in an image and populate it based on [subj_id, obj_id] = pred
I wonder how the subj_id/obj_id does not exceed the number of bbox in each image considering the number of categories in VG. Perhaps it has more to do with what im_to_first_rel & im_to_last_rel are from the load_graphs function.
I hope you can provide more clarity on the same. Thank you in advance.
I am trying to run your graph-rcnn on a custom scene-graph dataset. I had some questions regarding how
__getitem__
in dataloader creates the targets. I noticed that you create a 2D arrayobj_relations
with the shape of number of bboxes in an image and populate it based on[subj_id, obj_id] = pred
I wonder how the subj_id/obj_id does not exceed the number of bbox in each image considering the number of categories in VG. Perhaps it has more to do with whatim_to_first_rel & im_to_last_rel
are from theload_graphs
function. I hope you can provide more clarity on the same. Thank you in advance.