why sequences in provided train_parent.txt and DAVIS2016/train.txt are not the same?

Question

why sequences in provided train_parent.txt and DAVIS2016/train.txt are not the same?

nijkah opened this issue 6 years ago · comments

I guess parent model file provided from README is trained on your own sequences written in train_parent.txt.
Is there any purpose the provided model is trained on your own sequences?

Or is just train_parent.txt example of training sequences?

Kevis-Kokitsi Maninis · Answer 1 · Mon May 14 2018 20:45:47 GMT+0800 (China Standard Time)

Hi, this is the training set of DAVIS-2016. In the results reported in the paper, we used the training set to train the parent model, and the validation set to finetune on the first mask.

Sergi Caelles · Answer 2 · Mon May 14 2018 20:45:55 GMT+0800 (China Standard Time)

If you are referring to the train.txt in DAVIS/ImageSets/480p, the only difference with the train_parent.txt is that in the latter the / preceding the paths is removed. The sequences are the same in both files and they are the ones in the train set of DAVIS 2016. Those sequences are the ones used to train the parent model provided in the README. In this way, the model learns a notion of object saliency that improves the results in the final fine tunning. Let us know if you have any further question.