Question about GT #30
-
|
What a great job! I am deeply impressed with your work. I have one question about GT("instance_{train2015, test2015}.json). In the bounding box annotation("boxes_h" and "boxes_o"), there is slightly difference on same human and object like "boxes_h": [ Is this because it came from the detector, not the GT conversion? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hi @tiruss, The HICO-DET dataset was collected by Chao et al., I merely implemented some utilities. The annotation files are in fact converted from the original files. I don't know why the boxes are different for the same object instance, but my guess is that the annotations were collected class by class. As a result, the same bounding box for a different interaction class might be annotated by a different person, hence the minor difference. Fred. |
Beta Was this translation helpful? Give feedback.
Hi @tiruss,
The HICO-DET dataset was collected by Chao et al., I merely implemented some utilities. The annotation files are in fact converted from the original files.
I don't know why the boxes are different for the same object instance, but my guess is that the annotations were collected class by class. As a result, the same bounding box for a different interaction class might be annotated by a different person, hence the minor difference.
Fred.