I tested extd64 only got 5 fps

[lucasjinreal]

The model is tiny but speed is still not good though. It runs 0.4s per image, with some sample image from wider face.

Is that normal?

[yjyoo3312]

Thank you for the comment. The images in wider face are mostly very large (largely over 1200x1200), and in that case, most face detectors including ours run slowly, so I think it can be possible. Usually, the face detectors measure the speed using 320x320 or 640x640 images. For more information, please refer the FPS graph in Figure 6 in our paper.

Thank you.

[lucasjinreal]

@yjyoo3312 Beside the size of image, does faces num effect inference speed? I mostly test on dense face image.

[yjyoo3312]

@jinfagang In evaluation, yes (because of NMS), but in demo setting, the face number does not largely affect the inference speed.

[aniketvartak]

when i run the model it consumes ~3.5GB of memory. any idea why so?