VID is not good?

[Yanxudong23]

Thank you for your implement version for SiameseRPN, I recently recurrent Siam-RPN with zkisthebest's work, I used VID dataset to train, but the results are so bad, and I just now used your version with pretrained model on VID to test OTB, results also bad, I wonder why training with VID can't get good results even though VID is also a large dataset

[songdejia]

@Yanxudong23 Thanks for your reply. You can try to filter out some large object images. This is a serious problem we concerned about. We are now training it on Youtube bb, if we have any further progress, we will update this work.

[Yanxudong23]

Thanks for your reply. I see you have some snapshots such as test.git which has a good result, I want to know if the results comes from tracking with the pretrained model with VID or with the model trained by VID and Youtube bb.

[songdejia]

We train the model on a part of Youtube_bb, VID and other datasets, and selected a better result on vot2013, although many results are not good enough. Today I met wang, a post-doctor in Department of Engineering Science, Oxford, "siamese-FC" is from this group. He says, when it is traning, we must filter out large objects and only VID is Absolutely not enough.

[Yanxudong23]

Thank you~, look forward to your good news~.

[zzpustc]

@songdejia How to define large object in your code? Is that the ratio of the groundtruth to the image is greater than a certain threshold? Then how much the ratio is?