attn avg

[bmmi7]

Hi can you explain this line?

Easi3R/dust3r/cloud_opt/base_opt.py

Line 224 in 146a77d

att[0,0] = (att[0,1] + att[1,0])/2

[rover-xingyu]

Hi @bmmi7, the attention map att[0,0] reveals artifacts attributed to the register token, reflecting shortcut learning as noted in Vision Transformers Need Registers and Denoising Vision Transformers. We address this by averaging neighboring tokens.

[rover-xingyu]

BTW, you can easily see the phenomenon by commenting out this line of code.