PriorBox: the box is out of the image about ssd.pytorch

[AceCoooool]

And I am also confusing : in the decode function using the variance, but there is no "variance operation" in the prior_box.

[amdegroot]

Hi, I apologize for taking so long to get back to you. Finishing up school this week and then will be able to spend more time on this. I will get back to you on the first question ASAP, but I want to give you a good response and I don't have time at the moment. As for your second question, the decode function (which is only used at test time) includes the variance because the variance is encoded in the encode function at training time.

[jeong-tae]

Hi, is there a any answer about this issue? the output box is out of image!

[amdegroot]

@jeong-tae Can you please provide a specific example? I personally have not run into this issue.

[jeong-tae]

@amdegroot same issue with the @AceCoooool
the prior boxes has a out of image coords such as 0.8 0.8 0.5 1.0. is it normal in v2?
i got out of image boxes with my dataset, so i track the values. and i found out the boxes are little bit skewed to bottom right(label matched with skewed one, i dont know why). for example, label cxcy has 0.3 0.3, but after encode it returns 0.5 0.5. but w,h are quite similar bw encode decode. also gt and pred

[jeong-tae]

@amdegroot Here is more specific example for trace this. The truths loc was [0.34, 0.17, 0.74, 1.0] before the encode. To do encode, match first and then encode it. Matched prior box was [0.53, 0.32, 0.38, 0.76]. The result of encoded loc was [2.7, 8.85, -3.0, 1.13]. To ensure this process is right, i tried decode encoded loc. it returns [0.58, 0.13, 0.77, 1.07].

[jeong-tae]

Hello. I think this is only problem of my code. When i get the matches, i revised some codes and it may occur this issue(i didn't know this was modified on my code). Let me check. if still have problem, let you know. Sorry for the bothering you.