search

haomo-ai / MotionSeg3D

[IROS 2022] Efficient Spatial-Temporal Information Fusion for LiDAR-Based 3D Moving Object Segmentation
https://npucvr.github.io/MotionSeg3D/
GNU General Public License v3.0
240 stars 21 forks source link

MetaKernel annotation error #3

Open fengluodb opened 2 years ago

fengluodb commented 2 years ago

I found some annotation error in MetaKernel.

  1. In forward, param coord_data: num_batch, 3, H, W link. But it's not real, in SalsaNextWithMotionAttention, coor_data is current_range_image, it's channel is 5;
  2. In sample_data , :return: sample_output: num_batch, num_channel_in * kernel_size * kernel_size, H, W, link. I think the outout should be num_batch, num_channel * kernel_size * kernel_size, HxW
MaxChanger commented 2 years ago

Thank you for such meticulous coding review. I took a quick look and recalled that this part of the annotation may be from the original RangeDet, which has not been carefully checked and modified.

  1. It really needs to be corrected here, because in fact, any dimension can be passed in here, (x, y, z), or (x, y, z, i, r) https://github.com/haomo-ai/MotionSeg3D/blob/dc4c95fcdba2f0819d2bbc4a419f231c55e9c6f3/modules/SalsaNextWithMotionAttention.py#L136-L140

  2. I think you might be right, I look further into docs doc_mxnet and doc_torch_nn.Unfold, and confirm the tensor shape/size when debugging.

These will be fixed and updated soon.