search

google-research / ssl_detection

Semi-supervised learning for object detection
Apache License 2.0
408 stars 76 forks source link

About Cutout size #24

Open Chrisfsj2051 opened 3 years ago

Chrisfsj2051 commented 3 years ago

Hi!

I found that in your code, you set cutout_op as iaa.Cutout(nb_iterations=(1, 5), size=[0, 0.2], squared=True), here size=[0,0.2] means the size is either 0 or 0.2 ( under imgaug 0.4.0), which is different as what is described in paper.

Maybe we should change it to iaa.Cutout(nb_iterations=(1, 5), size=(0, 0.2), squared=True) ?

zizhaozhang commented 3 years ago

That is a great observation. We did not notice this difference during implementation. If you would like to compare the difference and let us know that will be great.

Chrisfsj2051 commented 3 years ago

That is a great observation. We did not notice this difference during implementation. If you would like to compare the difference and let us know that will be great.

In the process of reproducing the result of C+{GB}+Cutout under fully supervised manner, standard 1x setting in Table3, I tried to use cutout with binary {0%, 20%} and random [0%, 20%], here is the results:

method mAP
binary {0%,20%} 35.4
random [0%, 20%] 35.6

So I think it may make a small difference.