Some basic questions about the paper

[serkansulun]

Hi David,
I'm a masters student wanting to work on deep networks for tracking and I'm trying to implement your paper for Julia language. However I didn't understand some parts:

1- For ALOV300+, we have the ground-truth bounding box for the frames 1, 6, 11 and so on. We need 2 annotated frames so suppose we are using frames 6 and 11. What I assumed was, for a single forward run in the network, we crop twice the area around the GT box corresponding to frame 6, in frames 6 and 7 (which construct the inputs of the network), make a prediction for bounding box location in frame 7 (output), using this prediction like a new ground-truth, repeat for frames 7 and 8, all the way to frame 11 and calculate the loss between prediction and ground-truth boxes for frame 11. Is this correct?

2- How about data augmentation? On which frames do we take random crops? What about the ground-truth labels for these random crops? This is the part where I'm really lost. Same for imagenet dataset.

3- How does the network deal with varying size input (because ground-truth bounding box sizes aren't constant), since the fully connected layer sizes are fixed?

Appreciate your help.
Serkan

[serkansulun]

I might have discovered some things, looking at the code, please tell me if I'm wrong somewhere.

1- We are only using annotated frames, eg. only frame 6 and 11. Frame 6 becomes previous, frame 11 becomes current.

2- In augmentation we shift and scale the ground-truth box in the current frame, and crop twice the area around the shifted box, only to define a new synthetic search region. While getting labels (coordinates relative to the new search region), we still use the not-shifted ground-truth box coordinates.

Are these correct?

I'm still not clear about question 3.

Thanks

[davheld]

1-2) That is correct

3. After figuring out what part of the image we want to crop, we crop that region, and then we resize the cropped region to the fixed-size input for the network (227 x 227).

[serkansulun]

Thanks for the quick answer.

3. Do we do it by matching the shorter side to 227 and zero padding the remaining area? Do the ground-truth bounding box coordinates stay the same?
   Thanks

[davheld]

We do it by resizing both sides to 227x227. You can try the padding approach if you wish. The ground-truth bounding box coordinates are resized accordingly.

[serkansulun]

Hi again David,
To test my implementation, I want to use accuracy and robustness errors, instead of ranking. In your paper I couldn't find a direct reference to these metrics but I have some idea, so please let me know if I'm wrong somewhere.
1- Accuracy is simply intersection-over-union value, between ground-truth and predicted boxes.
2- When the accuracy becomes zero (no intersection), the tracker is reinitialized, meaning the ground-truth box is provided for the next frame. Or is it reinitialized n frames later?
3- Robustness error is number of reinitializations divided by number of frames.
4- Are these two measures averaged first over frames (in a single sequence) then over all sequences; or over all frames in all sequences at once?
Thanks in advance

[davheld]

To test my implementation, I want to use accuracy and robustness errors, instead of ranking. In your paper I couldn't find a direct reference to these metrics but I have some idea, so please let me know if I'm wrong somewhere.

See Table 1 http://davheld.github.io/GOTURN/GOTURN.pdf For the other questions - I'm not sure, sorry. You'd have to check with VOT or look at their code Best of luck!.

…

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#51 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEHoHJyjZwa-RXK85tZXfBX9qn4iMsbWks5sWw4JgaJpZM4NaKBW> .