why no encoder like detr2

[xudh1991]

I'm very sorry, although I have seen previous inquiries about this issue, I still want to confirm again because I think this position is somewhat unbelievable
As shown in the figure
Personal understanding: Before the Add operation, it corresponds to the position encoding in DETR, and after the Add operation, it corresponds to the transformer encoder in DETR.
But in your 3Dposition aware Features, there are no complex convolution operations, only an F, which is a flattening operation. That is to say, you completely abandoned the transformer encoder instead of using the 3D Position Encoder to replace this process.
The 3D Position Encoder only changes dimensions and does not involve additional learning. Should this be understood as such??
If it is completely discarded, have you ever tried adding a transformer encoder to improve the accuracy of the model