Open suzy0223 opened 5 years ago
Could you please read the paper in Section 3.5? " During training, the input image is resized to 640 × 640, and we randomly choose a number from {640, 576, 512, 384, 320} as the width and height to randomly crop patches. Finally, the cropped patches are further resized to 576 × 576."
So it is much bigger than other papers used. I found that the result in table1 used the report result in other paper directly(Resnet-srn)?
Yes, we directly use their results in the paper for comparison.
ok, thx
Yes, we directly use their results in the paper for comparison.
It is not fair to compared to other methods, you should report the results with 224 size. In addition, you use resnet-101 as your backbone, but other methods use VGG-16 as backbone in VOC dataset, it is also unfair.
Have your ever try 224 as the input size with your own code ? if did, can you share the comparison with other methods ? Because we can not completely reproduce your results
Have your ever try 224 as the input size with your own code ? if did, can you share the comparison with other methods ? Because we can not completely reproduce your results
have you ever run the code with 224 size? if you did,could you please tell me the result you got? thx.
in your paper, you said the size of the input image is 448, however, in main_coco.sh the crop_size is set to be 576. So could you please tell me what is the image size you use to complete the experiment.