search

songw-zju / Meta-RangeSeg

The official implementation of "Meta-RangeSeg: LiDAR Sequence Semantic Segmentation Using Multiple Feature Aggregation" (RA-L with IROS 2022)
MIT License
45 stars 6 forks source link

Question about Boundary Loss #2

Closed Carl12138aka closed 2 years ago

Carl12138aka commented 2 years ago

Nice work! Thank u for releasing this repo! I would like to ask 2 questions about BoundaryLoss~

  1. Does BoundaryLoss consume a lot of time in your experiments? In my experiments, computing the Boundaryloss consumes a lot of GPU memory and triples the training time.
  2. I don't use the range image method to represent the point cloud, but use the BEV two-dimensional representation, could the BoundaryLoss be used to train the network?

Thanks in advance for your reply!

songw-zju commented 2 years ago

Hi, @Carl12138aka, sorry for the late reply.

  1. In my experiments, the boundary loss takes a little memory and time consumption. This may be related to using range image as representation.
  2. I haven't tried applying boundary loss in bird's eye view. The boundary loss comes from the remote sensing image, and the point cloud under the bird's-eye view has more in common with the remote sensing image. I think it may be helpful, but not sure.
Carl12138aka commented 2 years ago

Thanks a lot for your reply! That helps me!

emilyemliyM commented 2 years ago

Nice work! Thank u for releasing this repo! I would like to ask 2 questions about BoundaryLoss~

  1. Does BoundaryLoss consume a lot of time in your experiments? In my experiments, computing the Boundaryloss consumes a lot of GPU memory and triples the training time.
  2. I don't use the range image method to represent the point cloud, but use the BEV two-dimensional representation, could the BoundaryLoss be used to train the network?

Thanks in advance for your reply!

Hi, could I know how to get the BEV two-dimensional representation from LiDAR data?

really thanks！！

songw-zju commented 2 years ago

Nice work! Thank u for releasing this repo! I would like to ask 2 questions about BoundaryLoss~

  1. Does BoundaryLoss consume a lot of time in your experiments? In my experiments, computing the Boundaryloss consumes a lot of GPU memory and triples the training time.
  2. I don't use the range image method to represent the point cloud, but use the BEV two-dimensional representation, could the BoundaryLoss be used to train the network?

Thanks in advance for your reply!

Hi, could I know how to get the BEV two-dimensional representation from LiDAR data?

really thanks！！

Hi, @mengshiyu0109 . For bev projection, you can refer to PointPillar(for object detection) and PolarNet(for semantic segmentation).