Results on only one of the dataset, e.g. VID.

[willtwr]

Has anyone tried to train with only one of the dataset (e.g. only VID) and evaluate the model?
It is much appreciated if anyone who did such experiment can share the result.
Thanks!

[xi-mao]

I have tried to train only using vid，but when I test ，I found the tracker cann't track the object at all。I don't know why。 I think it probably something wrong in my traning data。 But I generate traning data as the code completely。

[NineBall9]

I have trained only using vid, the result is 0.26, the difference is too large.

[xi-mao]

I have trained using vid and ytbb. The auc on otb100 is 0.54。 I also trained only using vid。 but i havn't test it。

[CXSunshine]

I have trained using vid and ytbb. The auc on otb100 is 0.54。 I also trained only using vid。 but i havn't test it。

How did you download the ytbb datasets?It is too big and need to google.

[willtwr]

Since it has more number of iterations when using 4 training sets compared to 1 or 2 of the datasets, perhaps training the tracker for more iterations might help?

[NineBall9]

I have trained using vid and ytbb. The auc on otb100 is 0.54。 I also trained only using vid。 but i havn't test it。
@xi-mao Could you test the model trained only on vid , and tell me the result on VOT2018? Thank you!

[zhanght021]

I have trained using vid and ytbb. The auc on otb100 is 0.54。 I also trained only using vid。 but i havn't test it。
@xi-mao Could you test the model trained only on vid , and tell me the result on VOT2018? Thank you!

@NineBall9 I have trained using vid coco and GOT-10k, the eao in VOT 2016 is 0.383, but eao in VOT2018 is only 0.28-0.30, the eao in VOT2019 is only only 0.245, the backbone i used is ### alexnet###

[cunjif]

Please, can tell me your hardware configs ? 8GPUS?

[lzx1413]

@zhanght021 researchmm/SiamDW#20 (comment)
pay attention to 3,4,5

[Ling-wei]

I trained SiamRPN++(resnet-50) for 20 epochs, using COCO、VID、DET, and only got EAO 0.31 for VOT2018 and 0.25 for VOT2019 . The best checkpoint comes checkpoint_10.pth, and the 10-20 Epochs training for ResNet&SiamRPN++ doesn't work.

[songdejia]

@Ling-wei So four datasets and enough iterations are necessary?

[XiaoMingDe]

+1，修改了bacnkone,并且加载res50，那么是否要修改epoch,以及是否要修改开始backbone_train_epoch(默认是10)，修改了backbone意味着要提前训练backbone咩

[charlesyann]

I trained SiamRPN++(resnet-50) for 20 epochs, using COCO、VID、DET, and only got EAO 0.31 for VOT2018 and 0.25 for VOT2019 . The best checkpoint comes checkpoint_10.pth, and the 10-20 Epochs training for ResNet&SiamRPN++ doesn't work.

get the same result with you, it seems youtube-bb dataset is necessarily needed. @Ling-wei

[NineBall9]

I have tried to train only using vid，but when I test ，I found the tracker cann't track the object at all。I don't know why。 I think it probably something wrong in my traning data。 But I generate traning data as the code completely。

@xi-mao ，how long it costs to train with data only VID, backbone resnet50, epoch 20 and 8 gpus?

[xi-mao]

@NineBall9 Traning SiamRPN with vid toke me one and half days using two GPUs

[willtwr]

Has anyone tried to use all 4 datasets to reproduce the results?

[gongbudaizhe]

Training with only VID won't work. You need at least VID + COCO to get reasonable results.
For now, the best results I can get using VID + COCO + YTB on OTB100 is 0.681 without searching hyperparameters. The most stable results are 0.676.

Using the same hyper-parameters with models at different epochs, the best EAO on VOT2018 is 0.317.