Open Dominic23331 opened 2 weeks ago
We plan to release the code within a few weeks. Please wait a little longer. For the 'Max Survival algorithm': We transform the shape into a defined grid format and then extract the max index. Subsequently, we apply a technique that changes everything except the max index to zero. This process is similar to max pooling, but the overall shape is maintained.
Did you use an attention mask here? I think each line should pay attention to its corresponding mask
What should I do if there are not enough points after threshold filtering in this step? If there is not a single point after passing the threshold, how should we proceed with the next steps?
If there are not enough points, use zero-padding. The code will be released later, and it will provide more details. Thank you.
What should I do if there are not enough points after threshold filtering in this step? If there is not a single point after passing the threshold, how should we proceed with the next steps?
Can you release this part of the code first?
May I ask what is the difference between using GFE module and directly using bev seg mask as the attention mask? @SehwanChoi0307
Using BEV features with a BEV segmentation mask as an attention mask can implicitly capture geometric information. However, this approach may require substantial computational resources and include unnecessary information. We propose GFE to explicitly capture the detailed geometric structure of each instance at the point level.
How to keep GFE module differentiable and efficiency for end to end training?
May I ask when the code can be released? One more thing, how is the Max Survival algorithm implemented in the GFE module? I use kernel to scan the mask, but the speed is too slow