Annotation format for bbox is not uniform in training set

[YiweiLi4]

Hello,

I found that most bbox annotations are ordered as [x1, y1, width, height], but some images seem to have the annotation like [x1, x2, y1, y2], like "HICO_train2015_00012054.jpg" below. How can I standardize these annotations?

Thanks for your time!

[rex-yue-wu]

Notice the same issue. A quick and dirty fix is to evaluate the quality of the two possible options, namely,

• IoU(bbox_xywh, keypoint_min_bbox)
• IoU(bbox_xyxy, keypoint_min_bbox) where bbox_xywh and bbox_xyxy are the two possible formats of the given bbox annotation, and keypoint_min_bbox is the minimum bounding box you may infer based on the given keypoints annotation.

Once both scores are computed, you can simply pick the one attaining better quality.