How to evaluate on YouTube-Objects and Long-Videos datasets?

[xyzskysea]

How to evaluate on YouTube-Objects and Long-Videos datasets?
and where to download corresponding dataset with groundtruth? could you provide the results of these two datasets?

[CODE4UVOS]

Please refer to our related project.

[xyzskysea]

i follow [The YouTube-Objects dataset can be downloaded from here and annotation can be found here.].

Do you use the dataset release v2.3? if possible, could you please provide the clean-up dataset and annotations directly? goole drive or baidu?

[CODE4UVOS]

The YouTube-Objects dataset has already been uploaded to GoogleDirve.

[xyzskysea]

In the YO2SEG, the Annotations are the groundtruth, mask folder is the results. Right?

[CODE4UVOS]

'Annotation' is GT (0,255), 'mask' is also GT (0,1).

[xyzskysea]

ok. i tried that. but found that the size of aeroplane0001/00001.png is 854x480. but in the results of HGPU it is 480x360. And i also found that in the AGNN results, the size is also 480x360. So which is the correct images and groundtruth?
it's weird!

[CODE4UVOS]

They are both correct. '480x360' is the original size and '854x480' is after we resized it.

[xyzskysea]

but when i tried to use PyDavis16EvalToolbox to evaluate. it mentions that the size is not consistent. so how can i do it to reproduce the result in your paper?

[CODE4UVOS]

The validation tool library is available at davis-matlab.

[chengguanjun]

@xyzskysea Could you test on YouTube-Objects dataset to achieve the results in the paper?

[qian507]

The YouTube-Objects dataset has already been uploaded to GoogleDirve.

Why does the YouTube-VOS data only contain val, but not train?