search

liulei01 / DRBox

A deep learning based algorithm to detect rotated object, for example, objects in remote sensing images
Other
421 stars 142 forks source link

question about your dataset #11

Open Zico2017 opened 6 years ago

Zico2017 commented 6 years ago

Did you evaluate the three models(DRBox, Faster R-CNN, SSD) with the totally same dataset? It's not so clear to me when I read your paper.

I mean Faster R-CNN and SSD only works for the horizontal bounding box, so you create a larger bounding box to cover the rotated bounding box on the same image? Then DRBox uses the smaller rotated bounding box, faster R-CNN uses the larger horizontal bounding box.

There is limited research on rotated bounding box in this domain, you provide a nice way to solve it, hope you will provide more continuous valuable work. (e.g. when there are many overlap objects, whether there are some conflicts in predicting bounding box for each of them, especially when we use non-maximum suppression approach to remove duplicate neighbor bounding box)

Thank you for your work!

Zico2017 commented 6 years ago

Another question:

on your paper, in (6a)(6b)(6c) formula, what the t_cx, t_cy...mean, ground truth, or something else? thank you in advance!