Getting landmark coordinates

[richliao]

Dear Ziwei,

Awesome and interesting work!

Do you mind shed some lights on what is purpose of the statement -get_orig_coordinate = @(p)((p+0.5)*224-repmat([offset(2),offset(1)]',[pipline.num_points,1]))/scale;?

I can not relate this to the paper, particularly (p+0.5)*224. I don't have matlab so I won't be able to debug the value but when I run pyCaffe those values of landmarks came out of stage 1 forward are very small, same as pseudo labels (all lower than 0.01). Any explanations will be greatly appreciated! Thanks.

[yysijie]

Dear richliao,
Thanks for your interest in our work.
We normalized landmarks to [-0.5, 0.5] in training process.
And we used images of 224*224 for training model.
This operation is just for projecting normalized landmarks to absolute coordinate frame.

[richliao]

Thanks for the explanation, Ziwei!

Can you further elaborate a bit how to convert from a vector (1 dimension) to x and y coordinates which are 2 dimension? Further, I don't see any input box, are you assuming the box is the full size of the image (224x224)? Thanks much.

[yysijie]

the meaning of a output vector is [x1, y1, x2, y2 ....]

Yes, and the matlab script would resize images to 224*224 before testing.

[liuziwei7]

@richliao Yes, the inputs to our Deep Fashion Alignment (DFA) are clothes bounding boxes. We treat this detection and cropping procedure as pre-processing and don't include it in this codebase. But you can definitely find bounding box annotations in the DeepFashion dataset.