andyyuan96 / pointclouddynamicvoxel

PointCloudDynamicVoxel

Why this project?

​ For 3D detection with point cloud, voxelization is a very common operation, and lots of 3D detection method use voxelization as first step to process point cloud, like VoxelNet, SECOND. But most of 3D object method use hard voxel(set max_points_in_voxel, and if the number of points in voxel exceed max_points_in_voxel, just throw the extra point).

​ In paper End-to-End Multi-View Fusion for 3D Object Detection in LiDAR Point Clouds(MVF), the author comes up with dynamic Voxel. Generally speaking, keep all the points in each voxel.Comparing to hard voxel, dynamic voxel uses less memory, and can keep all points.

​ I just implement dynamic voxel in the paper above according to my understand.

​ MVF gives an example of how to use the project.

How to use?

The project contains two parts, PointCloudVoxel and scatter, and both of them are written as a pytorch extension.

enviroment:

Tested on ubuntu16.04, cuda10.0/cuda9.2, python3.6/python3.7,  pytorch-1.1/pytorch-1.4

install:

cd PointCloudVoxel && python setup.py install
cd scatter && python setup.py install

test:

cd PointCloudVoxel/test && python test.py
cd scatter/test && python test.py

pointclouddynamicvoxel's Issues

dynamicVoxelBEVFaster(): incompatible function arguments.

Hello, I met the problem when training, but I had already compiled the PointCloudVoxel and scatter-max in my virtual env.

发生异常: TypeError
dynamicVoxelBEVFaster(): incompatible function arguments. The following argument types are supported:
1. (self: PointCloudVoxel.PointCloudVoxel, arg0: at::Tensor, arg1: at::Tensor, arg2: at::Tensor) -> None

Invoked with: <PointCloudVoxel.PointCloudVoxel object at 0x7fa832a8ed70>, tensor([[ 4.2830, 8.0850, -2.1385, 0.2700],
[ 3.6356, 17.1089, -1.9634, 0.2500],
[ 4.9427, 9.3864, -0.1725, 0.1500],
...,
[ 4.6178, 10.6555, -1.3861, 0.4700],
[ 4.3192, 10.4141, -1.9983, 0.3000],
[ 9.7151, 11.6478, -0.5648, 0.3600]]), tensor([[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.],
...,
[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.]]), tensor([0, 0, 0, ..., 0, 0, 0], dtype=torch.int32), tensor([[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
...,
[0, 0, 0],
[0, 0, 0],
[0, 0, 0]], dtype=torch.int32)
File "/home/yf/Programs/HVNet/local/pcdet/datasets/dataset.py", line 256, in prepare_data
self.hv_features[j].dynamicVoxelBEVFaster(points, bev_local_coordinate, bev_mapping_pv, bev_mapping_vf)
File "/home/yf/Programs/HVNet/local/pcdet/datasets/kitti/kitti_dataset.py", line 463, in getitem
example = self.prepare_data(input_dict=input_dict, has_label='annos' in info)
File "/home/yf/Programs/HVNet/local/tools/train_utils/train_utils.py", line 28, in train_one_epoch
batch = next(dataloader_iter)
File "/home/yf/Programs/HVNet/local/tools/train_utils/train_utils.py", line 82, in train_model
cur_epoch=cur_epoch
File "/home/yf/Programs/HVNet/local/tools/train.py", line 147, in main
max_ckpt_save_num=args.max_ckpt_save_num,
File "/home/yf/Programs/HVNet/local/tools/train.py", line 154, in
main()

About dimensions of points

Must the feature dimension of points be XYZ + intensity？Can there be more dimensions？

error: command 'x86_64-linux-gnu-gcc' failed with exit status 1

Hi! This bug occurred when I compiled the module "scatter". Have you ever met this problem?

In file included from /usr/local/lib/python3.6/dist-packages/torch/include/ATen/Tensor.h:3:0,
from /usr/local/lib/python3.6/dist-packages/torch/include/ATen/Context.h:4,
from /usr/local/lib/python3.6/dist-packages/torch/include/ATen/ATen.h:5,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/types.h:3,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/data/dataloader_options.h:4,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/data/dataloader/base.h:3,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/data/dataloader/stateful.h:3,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/data/dataloader.h:3,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/data.h:3,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/csrc/api/include/torch/all.h:4,
from /usr/local/lib/python3.6/dist-packages/torch/include/torch/extension.h:4,
from cuda/scatter_max.cpp:1:
/usr/local/lib/python3.6/dist-packages/torch/include/ATen/core/TensorBody.h:268:30: note: declared here
DeprecatedTypeProperties & type() const {
^~~~
error: command 'x86_64-linux-gnu-gcc' failed with exit status 1

performance evaluation

Hi,Did you evaluate it for any performance improvements?

Sactter function: error: no matching function for call to ‘collect_arguments(pybind11::tuple&)’

While trying to build scatter using python setup.py install i got erroe below. Kindly help

/home/user/anaconda3/envs/mvf/lib/python3.6/site-packages/torch/include/pybind11/cast.h:2108:44: error: no matching function for call to ‘collect_arguments(pybind11::object)’
return detail::collect_arguments(std::forward(args)...).call(derived().ptr());