Can't train SiamRPN

[ISosnovik]

Hello,
Thank you for your work!

I wanted to train SiamRPN.
I followed your training procedure.
I used the provided training config.
I used VID, DET, YTB, COCO for training
But it gives me only 0.53 on OTB2013 after tuning.

I found a strange moment
I checked the paths you write and the look like '/data/home/hopeng/data/ytb/crop271'
so 271 is the size of the dataset you used
In the config you have VID of size 271 but you provide only VID crop255
The same for DET and COCO
Their sizes to do not match the size of in the config
For my training I just changed for example paths from coco/crop271 to coco/crop511
and so on. And then trained. But the result is significantly worse than expected.

Could you provide the correct config?
or could you provide the correct datasets?

[davinca]

@ISosnovik I got same issue like you, at the end, the training loss only converge to 0.03 for cls loss and 0.15 for reg loss.

[sjtuchenye]

I get the same issue. When training SiamRPN on GOT10K dataset, the segLoss converge to 0.14 after 14epochs. And the GOT10K dataset downloaded from Baidu Disk is also of size 255.

[ISosnovik]

@JudasDie Could you help us with this issue?

[JudasDie]

@JudasDie Could you help us with this issue?

Hi, as for the data, you could use the provided files in our new code. As for the bugs, it may need some time for me to check and fix. You may try the new model recently.

[ISosnovik]

@JudasDie I checked the COCO dataset from your new repo. It is also of size 511.
Which size did you use to train your SiamRPN model?

I am not willing to try a new model because I need to reproduce your SiamRPN+

[JudasDie]

@JudasDie I checked the COCO dataset from your new repo. It is also of size 511.
Which size did you use to train your SiamRPN model?

I am not willing to try a new model because I need to reproduce your SiamRPN+

@JudasDie I checked the COCO dataset from your new repo. It is also of size 511.
Which size did you use to train your SiamRPN model?

I am not willing to try a new model because I need to reproduce your SiamRPN+

There is no influence as long as the image size is larger than 255 + shift pixels.

[ISosnovik]

@JudasDie I got it. But why do I get 0.53 instead of 0.67 on OTB with SiamRPN when I follow your isntructions?

[JudasDie]

@JudasDie I got it. But why do I get 0.53 instead of 0.67 on OTB with SiamRPN when I follow your isntructions?

As I said, I need time to check and retrain the code to find the bugs. Besides, I haven't met such a low result before.

[Sanshi6]

Thank you for your work. We encountered the same problem and were unable to train a model with a success rate exceeding 0.53. Have you found the bug?