How to generate co-occurrence matrix?

[IreneMahhy]

Hi! Thank you for sharing your work.
I generated a matrix myself for V-COCO dataset, and I wanted to compare to your method to see if it is correct, but I didn't see you mentioned the generation of co-occurrence matrix in your paper.
I also modefied code in mat_to_json.py, my thought is to compare every coordinates of different hoi-classes in an image, if two hoi has same h-o coordinates, they can be regarded as co-occur. But it turns out that there are totally no identical coordinates in different hoi-classes. I wondered it may be the annotations, like[207, 32, 426, 299] and [205, 32, 426, 305] are not identical though they are actually the same box. I modified the identical condition to add a tolerant bias, but my results are not so close to yours.
Can you help me about the issue or give some details about how you generate your matrix?

[Dong-JinKim]

Thank you interest in my paper.

As mentioned in the paper, we count the image level statistics in order to easily obtain the co-occurrence matrix.
Even though the bounding box coordinates are different if the actions are occurring in the same image, we consider them as co-occurring, and the co-occurrence matrix constructed with this method is shown to be helpful.

If you have any other questions please feel free to ask me.
Best regards.

[IreneMahhy]

Thank you interest in my paper.

As mentioned in the paper, we count the image level statistics in order to easily obtain the co-occurrence matrix.
Even though the bounding box coordinates are different if the actions are occurring in the same image, we consider them as co-occurring, and the co-occurrence matrix constructed with this method is shown to be helpful.

If you have any other questions please feel free to ask me.
Best regards.

Thank you for your reply!
I've successfully generated the co-occurrence matrix you offered.
But I've encountered another problem, about the algorithm you use to generate anchor action label sets. While reading your paper, I suggest that there are n clusters in label set D which divided by n mutual exclusive actions, each presents an independent label space. Other actions are assigned to these clusters according to their correlation with these anchor actions(one can be assigned to multiple groups).
Then I reviewed your code, for ground truth cluster label, you introduce a json file 'hoi2gid1.json' mapping hoi category to cluster label, and each instance are assigned to a fixed cluster label according to this file, for example, there is an instance which ground truth hoi label is '133', then its cluster label is '23'. Which makes me confused is, obviously 117 actions are divided into 31 clusters('0' represents negative samples I guess), each belongs to one specific cluster. But they are not strictly mutual-exclusive. If action A and B can't co-occur, but C co-occurs with A and B in different images, how can we decide which cluster it belongs to? I think there's a paradox in paper and actual code. And I'm not sure if I get it right......
I would be very grateful if you can help me with my problems!