The parameter of the tracking loss

[Haelles]

Thank you for your nice work!

Since the training code of QueryTrack is not released, I hope you can share the following training details with me:

1. What are the meaning of $\alpha_t$ and $\gamma$ in the formula(7) (the definition of the tracking loss)
2. Is there any document about your contrastive focal loss?

Thanks again.

[vealocia]

Thank you for your nice work!

Since the training code of QueryTrack is not released, I hope you can share the following training details with me:

1. What are the meaning of $\alpha_t$ and $\gamma$ in the formula(7) (the definition of the tracking loss)
2. Is there any document about your contrastive focal loss?

Thanks again.

$\alpha_t$ and $\gamma$ is the same as $\alpha$ and $\gamma$ in focal loss.

In form, our contrastive focal loss is a focal loss with softmax function (Eq.5-Eq.7).

[Haelles]

Thank you for your nice work!
Since the training code of QueryTrack is not released, I hope you can share the following training details with me:

1. What are the meaning of $\alpha_t$ and $\gamma$ in the formula(7) (the definition of the tracking loss)
2. Is there any document about your contrastive focal loss?

Thanks again.

$\alpha_t$ and $\gamma$ is the same as $\alpha$ and $\gamma$ in focal loss.

In form, our contrastive focal loss is a focal loss with softmax function (Eq.5-Eq.7).

Thank you!
By the way, I have another question. During the calculation of losses['loss_cls'] in class DIIHead, the value of label_weight is always 1:

label_weights[pos_inds] = pos_weight
label_weights[neg_inds] = 1.0

However, when calculating avg_factor, only positive samples are included. Could you please tell me why? Thanks a lot!