Vram requirements

[maclory]

Vram requirements

[Steve-Tod]

In our experiments, we crop the images based on landmark annotations following previous works. First, we find a rectangular bounding box that can cover the landmarks tightly. Then we crop this bounding box to a square box with length equal to the short side of the rectangular box centered at the same point. Finally, we crop the image with the square box expanded by a scale factor of 1.3 and resize it to 128X128 by bicubic.
I think a large difference in cropping strategies might affect the performance while small differences might have sight effects since the alignment network can help localize the faces.

[Steve-Tod]

Hi, are you running the test or training?

[camjac251]

I'm attempting to test based on the pretrained models

[Steve-Tod]

We did our test with 16X16 input images and batchsize 1 and it takes only 1373MB GPU memory. Maybe you are using too large image or your GPU memory is already used up?

[camjac251]

I was using images that were about 100x100 pixels. I attempted to downscale the image to 16x16 and use that as an input but the result didn't look that good.

[Steve-Tod]

The unsatisfactory results might be caused by different preprocessing. In our experiments, images are cropped in a uniform way based on the detected face landmarks.

[DarkerDreamer]

The unsatisfactory results might be caused by different preprocessing. In our experiments, images are cropped in a uniform way based on the detected face landmarks.

Hi, I wonder what kind of preprocessing method is applied? Does different cropping method lead to different results?