May I ask how to understand the format of the train_set.txt? about goturn-tensorflow

[tangyuhao]

Bounding box is in the form of <x1, y1, x2, y2>. In your case, it means if the second frame has size
(h, w) #h is height and w is width
the target should be in the bounding box with top left corner (y1 * h, x1 * w) and bottom right corner (y2 * h, x2 * w)

And I suppose all first frames have bounding box of <0.25,0.25,0.75,0.75> so I do not need to input this to the network. The only label is the position of the bonding box in the next frame.

[tangyuhao]

However the network output is 10x of the original bbx. Thus it always outputs values ranging from 0 to 10. Sometimes exceeding it though.

[liyang-7]

Thank you very much! The CaffeNet resize the crop image into 227227, the input of the network is 227227, no matter the size(w, h) of the crop image. So the output of the network need to multiply w and h in order to regain the size of the crop image.
Because of your code, I understand the detail of GOTURN.
Extremely grateful!

[tangyuhao]

You are welcome~

[yihongXU]

Hi,
May I ask you why sometimes they have negative values ?
Evan in your test sest examples, I found some negative values
test_example/target/000020.jpg,test_example/searching/000020.jpg,-0.13129449838187718,0.29418666666666665,0.5079611650485438,0.95328

thanks

[tangyuhao]

because the actual object is outside the search area. I did this to test whether is have some ability to guess the actual position of an object if some part of it is not inside the searching area.

[liyang-7]

Hi, may I ask you the tensorflow version in this code?
I download the pretrained model from: GOTURN_MODEL.
Uncompress the checkpoints folder, then I got two items: checkpoint and checkpoint.ckpt-1.data-00000-of-00001
I got the error: Unsuccessful TensorSliceReader constructor: Failed to find any matching files.
I used the tf (1.1.0)
It seems that lack of the model.ckpt.index in the checkpoints folder.
Did you used the earlier tf version?

[tangyuhao]

I used 1.0.0, but 1.1.0 should also work fine. I have added the checkpoint index file. It should fix the bug.

[mrgloom]

How Bounding box is in the form of <x1, y1, x2, y2>, usually from 0 to 1, but exceeding this range is also fine. bbox encoding related to encoding proposed in paper?

Is encoding in section 4.2 Learning Motion Smoothness is only used at train time for augmentation?

[linhandev]

so does this mean when i use the output of the network to draw a bb, fc4/10 should be my x1,y1,x2,y2 and then use the above x1w , y1h , x2w , y2h as a pixel wise output