about network input size about cornernet-lite HOT 5 CLOSED

Following the default settings of YOLOv3, we evaluate
YOLOv3 at 3 single image scales (320, 416 and 608). Similarly, we also evaluate CornerNet-Squeeze at different single scales (0.5, 0.6, 0.7, 0.8, 0.9, 1). CornerNet-Squeeze
achieves a better accuracy and efficiency (34.4% at 30ms)
trade-off than YOLOv3 (32.4% at 39ms).

I think this comparison is unfair.
CornerNet-Squeeze scale 1.0 is the original image size or large than that.

from cornernet-lite.

Maybe this is an innovation.

from cornernet-lite.

In CornerNet-Squeeze, we follow the inference procedure in CornerNet, where we use images at the original resolutions as the input to the network.

from cornernet-lite.

OK @heilaw . Thank you.
If input the original image to yolov3, I think AP will be improved.
For CornerNet, the inference time will increase if the image resolution is large.

from cornernet-lite.

@xiaozhuka @heilaw
Have you test the forward of CornerNet-Squeeze is 30ms per image? But I test the result ,
CornerNet_Saccade T4:420ms 1080Ti:205ms
CornerNet_Squeeze T4:51ms 1080Ti:43ms

what is the condition of 30ms?
Thank you.

from cornernet-lite.