Why does box_w only divide by strides [-1]

[fengxiaoming520]

Hello, I would like to ask you a question: In the multi_gt_creator function of the tools.py file, why does box_w only divide by strides [-1] before calculating iou, does this not always cause iou to be calculated on the minimum or maximum downsampling multiple? Instead of calculating at all downsampling multiples.

box_ws = box_w / strides[-1] box_hs = box_h / strides[-1]

[yjh0410]

It is because that my all anchor boxes have been divided by 32(the largest stride). Therefore, before computing IoU, I have to divide the groundtruth box by 32. Even though some anchor boxes correspond to stride =8 or 16, see the formulations:                 tw = np.log(box_ws / p_w)                 th = np.log(box_hs / p_h)

It eliminates the effects of factor stride meaning that we actually needn't to divide the anchor boxes and groundtruth bxoes by stride and just omit strides. If I didn't divide my all anchor boxes by 32, I didn't need to divide the box_w and box_h by 32, either.

------------------ 原始邮件 ------------------ 发件人: "fengxiaoming520"<[email protected]>; 发送时间: 2020年3月21日(星期六) 晚上10:45 收件人: "yjh0410/pytorch-yolo-v2-v3"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: [yjh0410/pytorch-yolo-v2-v3] Why does box_w only divide by strides [-1] (#25)

Hello, I would like to ask you a question: In the multi_gt_creator function of the tools.py file, why does box_w only divide by strides [-1] before calculating iou, does this not always cause iou to be calculated on the minimum or maximum downsampling multiple? Instead of calculating at all downsampling multiples.

box_ws = box_w / strides[-1]
box_hs = box_h / strides[-1]

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[fengxiaoming520]

Let ground truth find the anchor box that suits it, and then use the anchor box to decide which downsampling multiple it belongs to. Do you mean that?

[yjh0410]

Yes. In my configuration, there are 9 anchor boxes where the first three anchor boxes correspond to stride=8, the middle three ones correspond to stride=16, and the last three ones correspond to stride=32. If the one of first three anchor boxes is chosen，then the groundtruth is allocated to stride=8 feature map.

------------------ 原始邮件 ------------------ 发件人: "fengxiaoming520"<[email protected]>; 发送时间: 2020年3月21日(星期六) 晚上11:11 收件人: "yjh0410/pytorch-yolo-v2-v3"<[email protected]>; 抄送: "ら ．Secret"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [yjh0410/pytorch-yolo-v2-v3] Why does box_w only divide by strides [-1] (#25)

Let ground truth find the anchor box that suits it, and then use the anchor box to decide which downsampling multiple it belongs to. Do you mean that?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[fengxiaoming520]

I see, thank you for the explanation.^v^

[yjh0410]

You are welcome ！！

------------------ 原始邮件 ------------------ 发件人: "fengxiaoming520"<[email protected]>; 发送时间: 2020年3月21日(星期六) 晚上11:23 收件人: "yjh0410/pytorch-yolo-v2-v3"<[email protected]>; 抄送: "ら ．Secret"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [yjh0410/pytorch-yolo-v2-v3] Why does box_w only divide by strides [-1] (#25)

I see, thank you for the explanation.^v^

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[Haoyue-code]

您好，请问我在运行demo的时候出现这个问题怎么解决呢