visinf / veto Goto Github PK

Vision Relation Transformer for Unbiased Scene Graph Generation (ICCV 2023)

License: Apache License 2.0

Python 15.46% C 0.10% C++ 0.19% Cuda 1.21% Shell 0.01% Jupyter Notebook 83.02%

veto's Issues

About the graph constraint in the evaluation of MEET.

Under the graph constraint, the evaluator is supposed to take only one prediction for each subject-object pair.
However, in the implement of MEET, you just concatenates the pred_rel_scores and pred_rel_labels of each group. That means the evaluator will include 5 prediction for each subject-object pair, which might result in an unfair comparison.
Have you noticed this in your experiments?

veto/pysgg/modeling/roi_heads/relation_head/inference.py

Lines 376 to 397 in 832ba0e

 for scores in triple_scores_ensemble: 

 total_length += len(scores) 

 incr_length.append(len(scores)) 

 triple_scores_final = torch.zeros(total_length).cuda() 

 rel_pair_idx_final = torch.zeros(total_length, 2).cuda() 

 rel_class_prob_final = torch.zeros(total_length, num_rel_classes).cuda() 

 rel_labels_final = torch.zeros(total_length, dtype=torch.int64).cuda() 

 start = 0 

 for i, scores in enumerate(triple_scores_ensemble): 

 triple_scores_final[start:start+incr_length[i]] = scores 

 rel_pair_idx_final[start:start + incr_length[i]] = rel_pair_idx_ensemble[i] 

 rel_class_prob_final[start:start + incr_length[i], chosen_labels_incr[i]] = rel_class_prob_ensemble[i] 

 rel_labels_final[start:start + incr_length[i]] = rel_labels_ensemble[i] 

 start += incr_length[i] 

 _, sorting_idx = torch.sort(triple_scores_final.view(-1), dim=0, descending=True) 

 rel_pair_idx_ensemble = rel_pair_idx_final[sorting_idx] 

 rel_class_prob_ensemble = rel_class_prob_final[sorting_idx] 

 rel_labels_ensemble = rel_labels_final[sorting_idx] 

 boxlist.add_field('rel_pair_idxs', rel_pair_idx_ensemble) # (#rel, 2) 

 boxlist.add_field('pred_rel_scores', rel_class_prob_ensemble) # (#rel, #rel_class) 

 boxlist.add_field('pred_rel_labels', rel_labels_ensemble) # (#rel, ) 

 results.append(boxlist)

iam impressed this project i want to run the code for doing my final project .

so when i see the codes some of the init files there is no code .can you give the reason .
can this code run on cpu version

Have you encountered any computing power issues?

I am currently using Nvidia A100 GPU, cuda is using 10.2, and pytorch is also using the corresponding GPU version of 10.2. However, I have encountered the following issue. I see that you are also using Nvidia A100 GPU, cuda is 10.1. Do you have this issue

About the depth images

As mentioned in the article, you built the depth dataset with the AdelaiDepth tool, and the depth dataset is not currently in the repository, so the code is not available at this time, will you share the depth dataset you built later? Looking forward to your reply

i cant find the file visual_info.json

can you send to me [email protected]

about the inference.

Hi, I have some thoughts on your approach:
It seems that you ensemble the results of different models, which may cause an object pair to have more than one predicted relationship during the test, which does not meet the current unlimited sgg rules. As far as I know, all the methods you compared are with constraint, that is, only the top1 result of the current object pair is considered to be the predicate of the last evaluation.

Excuse me, I can't find the file. Can you provide the file？

"depth_img_dir": "/visinf/home/gsudhakaran/scene_graphs/Depth-VRD/VG_depth_raw_full", #"/data/user/dataset/scene_graph/images/VG/vg_depth_1024/VG_100K", #"/path/home/user/scene_graphs/Depth-VRD/VG_depth_raw_full", #,

No depth map files found

visinf / veto Goto Github PK

veto's Issues

About the graph constraint in the evaluation of MEET.

iam impressed this project i want to run the code for doing my final project .

Have you encountered any computing power issues?

About the depth images

i cant find the file visual_info.json

about the inference.

Excuse me, I can't find the file. Can you provide the file？

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

Jobs

	for scores in triple_scores_ensemble:
	total_length += len(scores)
	incr_length.append(len(scores))
	triple_scores_final = torch.zeros(total_length).cuda()
	rel_pair_idx_final = torch.zeros(total_length, 2).cuda()
	rel_class_prob_final = torch.zeros(total_length, num_rel_classes).cuda()
	rel_labels_final = torch.zeros(total_length, dtype=torch.int64).cuda()
	start = 0
	for i, scores in enumerate(triple_scores_ensemble):
	triple_scores_final[start:start+incr_length[i]] = scores
	rel_pair_idx_final[start:start + incr_length[i]] = rel_pair_idx_ensemble[i]
	rel_class_prob_final[start:start + incr_length[i], chosen_labels_incr[i]] = rel_class_prob_ensemble[i]
	rel_labels_final[start:start + incr_length[i]] = rel_labels_ensemble[i]
	start += incr_length[i]
	_, sorting_idx = torch.sort(triple_scores_final.view(-1), dim=0, descending=True)
	rel_pair_idx_ensemble = rel_pair_idx_final[sorting_idx]
	rel_class_prob_ensemble = rel_class_prob_final[sorting_idx]
	rel_labels_ensemble = rel_labels_final[sorting_idx]
	boxlist.add_field('rel_pair_idxs', rel_pair_idx_ensemble) # (#rel, 2)
	boxlist.add_field('pred_rel_scores', rel_class_prob_ensemble) # (#rel, #rel_class)
	boxlist.add_field('pred_rel_labels', rel_labels_ensemble) # (#rel, )
	results.append(boxlist)