Issues about Positional Embedding and Reference Point

[tae-mo]

Hi, thanks for sharing your wonderful work.

I got a question in here,

ConditionalDETR/models/transformer.py

Line 33 in ead865c

def gen_sineembed_for_position(pos_tensor):

which embedes positional information in the query_pos.

however, I don't understand the reason why does 2*(dim_t//2) has to be devided by 128, instead of the actual dimension pos_tensor has (e.g., 256 by default).

ConditionalDETR/models/transformer.py

Line 38 in ead865c

dim_t = 10000 ** (2 * (dim_t // 2) / 128)

Is it works correctly even dim_t is divided by 128?

I would appreciate to be corrected !

And another question is,
when we do the calculation of the equation (1) in the paper,

ConditionalDETR/models/conditional_detr.py

Line 89 in ead865c

tmp[..., :2] += reference_before_sigmoid

can I understand that the model would learn "offsets" from the corresponding reference points?
what is precise role of the reference points?

Thank you!

[Run542968]

Hi, for question (1), why does 2*(dim_t//2) has to be devided by 128, since the position embedding performs on both the x and y direction, then concat.