Is it necessary to replace 64 static anchors with 6 learnable queries?

[ZhanjunMei]

Hi,

Great work!

I noticed that your implementation replaces the original MTR's 64 static anchors with 6 learnable random initialization queries.

I tried applying Annealed WTA directly to the original MTR (with static anchors), but the training failed to converge. I was wondering if the switch to learnable queries is strictly necessary for convergence? Did you observe similar instability with static anchors during your development process?

Thanks.