Non Normalized coordinates

[vaishnavm217]

Hi,
I am working on face detection using your ssd model. One thing I noticed that my model was giving zero loss, i.e all the ious were zero. After long inspection

torchcv/examples/ssd/train.py

Line 63 in 6291f3e

boxes, labels = box_coder.encode(boxes, labels)

this line passes the boxes coordinates as it is. It should be normalized to get correct iou

[tfygg]

Yeah,
i also noticed that you tested the VOC dataset by the image size of 300*300.

[ljtruong]

@vaishnavm217 to fix this do we just do something like this? to normalise the bounding boxes.

        w,h = image.size
        boxes /= torch.Tensor([w,h,w,h]).expand_as(boxes)

[vaishnavm217]

That's more appropriate. But assuming a square image as input, you can just divide it by any one of the dimension.

…

On Sun, 15 Jul 2018, 10:22 Leon, ***@***.***> wrote: @vaishnavm217 <https://github.com/vaishnavm217> to fix this do we just do something like this? to normalise the bounding boxes. w,h = image.size boxes /= torch.Tensor([w,h,w,h]).expand_as(boxes) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#30 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQy9UHw_G5TtJTRyZOSs4DGpGYW8B_O4ks5uGsqmgaJpZM4UlnKK> .

[ljtruong]

@vaishnavm217 thank you for the confirmation. much appreciated. I'm testing out the train now and it's working well.

Just looking ahead. When I run it through prediction or evaluation, where would the best place to denormalise the labels?.