graph-rcnn.pytorch icon indicating copy to clipboard operation
graph-rcnn.pytorch copied to clipboard

[ECCV 2018] Official code for "Graph R-CNN for Scene Graph Generation"

Results 45 graph-rcnn.pytorch issues
Sort by recently updated
recently updated
newest added

Thanks for this great repository. In the ECCV paper, you used a VGG backbone. Could you please share its checkpoints? I understand you have provided a better ResNet backbone already,...

help wanted

When using relation heads (lib\scene_parser\rcnn\modeling\relation_heads\relation_heads.py) to predict relations with **ground truth** bounding box in training phrase, I got a weird error in [here](https://github.com/jwyang/graph-rcnn.pytorch/blob/master/lib/scene_parser/rcnn/modeling/relation_heads/relation_heads.py#L145) ``` File "/home/graph-rcnn.pytorch-master/lib/scene_parser/rcnn/structures/boxlist_ops.py", line 123, in cat_boxlist...

Thank you for a nice work! I found that there is pos_encoding argument in class Relationshipness (relationshipness.py) I guess it is to regress each relationshipness with considering geometric location of...

When training, I got loss "nan". What's the problem??

#68 `Traceback (most recent call last): File "main.py", line 14, in from lib.model import build_model File "/nethome/hluo76/GQA/graph-rcnn.pytorch/lib/model.py", line 9, in from .scene_parser.parser import build_scene_parser File "/nethome/hluo76/GQA/graph-rcnn.pytorch/lib/scene_parser/parser.py", line 9, in from...

Hi, When I evaluate using this chk: sg_baseline_joint_ckpt.pth. The result is not good. I use this config file: sgg_res101_joint.yaml, the parameters are set as: MODEL: WEIGHT_IMG: "catalog://ImageNetPretrained/MSRA/R-101" WEIGHT_DET: "sg_baseline_joint_ckpt.pth" RELATION_ON:...

Dear author, I feel confused when seeing this results. Does it mean that those complex models using contextualization cannot even beat the simple statistical freq model?

SCENE_PAESER_DICT instead of SCENE_PARSER_DICT in line 18 and 27

Great work & repo! I was wondering if there's a prediction pipeline available for this framework that can be applied to arbitrary images (non-VG) that do not come with ground-truth...

When I test with full VG dataset, one error appears: Traceback (most recent call last): File "main.py", line 129, in main() File "main.py", line 126, in main test(cfg, args) File...