If "DetrTransformerDecoderLayer" has been used in mapqr, is the "self_attn_mask" for one2many strategy works in self attention module?

HXMap / MapQR

[ECCV 2024] This is the official implementation of MapQR, an end-to-end method with an emphasis on enhancing query capabilities for constructing online vectorized maps.

https://arxiv.org/abs/2402.17430

MIT License

164 stars 8 forks source link

If "DetrTransformerDecoderLayer" has been used in mapqr, is the "self_attn_mask" for one2many strategy works in self attention module? #12

Closed TommyAnqi closed 1 month ago

fishmarch commented 2 months ago

Thanks for pointing out this bug. I did not notice this before, I will further check and fix this recently. The performance should be better after fixing this bug.

fishmarch commented 1 month ago

I upload a version to fix this bug. But I get similar perfomance with this attn_mask, it gets mAP1 = 42.9 and mAP2 = 66.6.