SuperZ-Liu / PolarBEV

The offical code of PolarBEV (CoRL2022).
MIT License
55 stars 3 forks source link

When can we see the code of PolarBEV? #1

Closed xuxiaoxxxx closed 2 years ago

SuperZ-Liu commented 2 years ago

No description provided. Thanks for your interest at our work. The code will be released at Sep.

xuxiaoxxxx commented 2 years ago

Thank you for your reply

xuxiaoxxxx commented 2 years ago

I'm very interested in this paper, and after I read this paper I have some questions. First, how can we iterate to update query embedding q(mentioned in 3.3)? Second, after we get the image features, how can we use it? concatenate with previous one? I would appreciate it if you could draw a simple network framework, I think it will help me understand this paper better. image

SuperZ-Liu commented 2 years ago

I'm very interested in this paper, and after I read this paper I have some questions. First, how can we iterate to update query embedding q(mentioned in 3.3)? Second, after we get the image features, how can we use it? concatenate with previous one? I would appreciate it if you could draw a simple network framework, I think it will help me understand this paper better. image

Sorry for the late reply. The query embedding $q$ are updated by sampling the 2D image features with the 2D-to-3D Feature Transformation. As for the second question, the image features are used to form the BEV representation and estimate the height of each query.

ghost commented 2 years ago

HI, I also interested in your work. According to your description, I draw a simple network framework. Can you help me check if it is correct? image

SuperZ-Liu commented 2 years ago

HI, I also interested in your work. According to your description, I draw a simple network framework. Can you help me check if it is correct? image

Yes, it is almost right except for missing the Ring Convolution after the transformation and iterating the whole process several times.

ghost commented 2 years ago

I get it, thank you for your reply!