how to deal with machine-generated labels of openimages ?

[tangsipeng]

how to deal with machine-generated labels of openimages ?

[LigZhong]

The images I downloaded does not match the label files either. Could anyone give a hit which version of data I should download.

[tangsipeng]

and i found many missing labels，such like 88fb29e99f00706c，i draw all labels that human‘s related.

[tangsipeng]

The images I downloaded does not match the label files either. Could anyone give a hit which version of data I should download.

as the paper says ,2019 release of OpenImages

[LigZhong]

@tangsipeng Yes, I downloaded the version as indicated but found a image size mismatch error. And the classification accuracy curve is also wired.

[xingyizhou]

Hi, Sorry for the delayed response. I am not aware of the machine-generated label issue in OpenImages. It is likely a version mismatch --- as mentioned above, we used the Challenge-2019 release. I am aware of the size mismatch issue, but don't have a good solution (even though I have tried re-computing the image size when generating a new json file). I just simply comment out the assert when checking the size, it doesn't affect the training at all.