search

yifanlu0227 / HEAL

[ICLR2024] HEAL: An Extensible Framework for Open Heterogeneous Collaborative Perception ➡️ All You Need for Multi-Modality Collaborative Perception!
Other
147 stars 9 forks source link

About aligners #12

Closed xhjy2020 closed 4 months ago

xhjy2020 commented 4 months ago

Hi, yifan. I noticed that the encoder in your code consists of three parts: encoder, backbone and aligner. The aligner does not seem to change the dimension of the feature map. So, what is its role in local training? Is it just used to increase the number of trainable parameters? Would removing it and just changing the encoder and bakbone parameters make a big difference to the alignment? Looking forward to your reply!

yifanlu0227 commented 4 months ago

No. Due to some of my previous setups, I put an aligner here. In the current framework, it is just a trainable network (added some network parameters, modified the number of channels for some feature maps, not necessary). You can remove it alive to set it to nn.identity

xhjy2020 commented 4 months ago

I understand, thank you very much!