About results on VOT2016

[Evelyn5589]

Hello ,DM

I have run the experienment on VOT2016 for several hours. And the EAO is 0.3581 on baseline , lower than 0.375 in the paper . I doubt whether the param causes the difference ?

[martin-danelljan]

Hi. We are currently investigating this issue. As mentioned, the results can vary somewhat on different computers. Are you using the mexResize function?

[Evelyn5589]

Hi, Martin Danelljan.Sorry to disturb you again. I have a question ..Why the results can vary somewhat on different computers.? I have run it on GeForce GTX 1080 and Tesla K40m. The results are different. On the other hand, does the result depend on whether it runs on CPU or GPU ?

[martin-danelljan]

Hi, that is no problem.

We have not published the full GPU version of the ECO code yet. I hope to get it ready before summer. But differences on different machines probably mostly depend on that the compilations of the underlying libraries for e.g. HOG, resizing, deep features etc. are slightly different. This is the case for any tracker that is not pure Matlab code. In fact, even the results for those trackers can vary.

[Evelyn5589]

Hi , Martin Danelljan I have another question. About DSST, you have considered the scale and translation filter. I think it can be implemented with the angle rotation filter. For many annotations in VOT2016/17 are all rotated . This is my thought . Sorry to disturb you again...~~. ^_^

发自我的 iPhone

在 2017年5月18日，下午8:48，Martin Danelljan notifications@github.com 写道：

Hi, that is no problem.

We have not published the full GPU version of the ECO code yet. I hope to get it ready before summer. But differences on different machines probably mostly depend on that the compilations of the underlying libraries for e.g. HOG, resizing, deep features etc. are slightly different. This is the case for any tracker that is not pure Matlab code. In fact, even the results for those trackers can vary.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[martin-danelljan]

Yes, you are welcome to try! We have played around with rotation filters long in the past, but never got it to work really stable.

[Evelyn5589]

I have tried to change the output layer of CNN.But I found that when I change the layer and down sample factor to control the cell size.Sometimes it goes wrong,such as, Output layer=[ 1 7 14]. Down sample factor=[ 4 2 1]. So, is there any restriction to these parameters ?

发自我的 iPhone

在 2017年5月19日，下午5:41，Martin Danelljan notifications@github.com 写道：

Yes, you are welcome to try! We have played around with rotation filters long in the past, but never got it to work really stable.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[martin-danelljan]

Did you remember to set the compressed dimensionality to something? It must contain the same number of elements as layers. For example: cnn_params.compressed_dim = [16 32 64];

[Evelyn5589]

Hello.Dear DM, I don't know why you have referred that the CNN feature can only be from the different layer of the same net .I think that it can also be from different nets. It's a little. strange.

发自我的 iPhone

在 2017年6月3日，上午12:32，Martin Danelljan notifications@github.com 写道：

Did you remember to set the compressed dimensionality to something? It must contain the same number of elements as layers. For example: cnn_params.compressed_dim = [16 32 64];

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[martin-danelljan]

Of course multiple nets can be used. And the tracker itself supports it. The only problem is that the init_features.m has to be rewritten slightly to support this. But that is pretty easy.

[Evelyn5589]

Yes ,I found that when the multi-net was used, the compressed_dim of the two different net must be same ,or there would be an error. And it shows that the result isn't good if the multi-net is used.

发自我的 iPhone

在 2017年6月3日，下午10:18，Martin Danelljan notifications@github.com 写道：

Of course multiple nets can be used. And the tracker itself supports it. The only problem is that the init_features.m has to be rewritten slightly to support this. But that is pretty easy.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[Ouya-Bytes]

＠Evelyn5589 How can I contact you if there are some questions I need your help?Please..