Attempting to understand the denoising.

[Alexandre-SCHOEPP]

Greetings,

I am attempting to understand the steps you took in your denoising.py (for both versions 1 & 2). I have a few questions:

While adding some noise to the bouding boxes, you have the following operation:

rand_sign = torch.randint_like(input_query_bbox, 0, 2) * 2.0 - 1.0
rand_part = torch.rand_like(input_query_bbox)
rand_part = (rand_part + 1.0) * negative_gt_mask + rand_part * (1 - negative_gt_mask)

As far as I can tell, the final operation is equivalent to running

rand_part = rand_part + negative_gt_mask

Since

$$ (r + 1) \times n + r \times (1-n) = (r \times n + n) + (r - r \times n) = r + n $$

Is there something I missed ? Are you using floating point inaccuracy for non-linearity ?