CenterPoint for Argoverse Data

This repository is a simplified version of the original CenterPoint repo, with far less code, designed purely for inference (not for training).

Installation

Environment:

Python 3.8.3

Dependencies

SparseConv (build locally)
DCN (build locally), produce deform_conv_cuda.cpython-38-x86_64-linux-gnu.so
Iou3dNMS (build locally), produce iou3d_nms_cuda.cpython-38-x86_64-linux-gnu.so
Pytorch 1.7.1 (check with python -c "import torch; print(torch.__version__)")
argoverse-api
CUDA 11.0 (check with python -c "import torch; print(torch.version.cuda)")

If you wish to run viz_aggregated_sweeps.py, you must run: Mayavi Environment: https://github.com/mne-tools/mne-python/blob/master/environment.yml

Bug Fixed addressed

pytorch/pytorch#29642

use torch::RegisterOperators

Not a problem if you use latest Pytorch?

nvcc fatal : Unknown option '-Wall' traveller59/spconv#69 CUDACXX=/usr/local/cuda/bin/nvcc python setup.py bdist_wheel pip install * --force-reinstall

https://pytorch.org/get-started/previous-versions/

RuntimeError: /nethome/jlambert30/spconv/src/spconv/indice.cu 274 cuda execution failed with error 98 invalid device function prepareSubMGridKernel failed traveller59/spconv#34 Make sure you use the same CUDA version for all installations (set CUDA_HOME before building anything)

Deformable Convolution

Added here: https://github.com/pytorch/vision/pull/1586/files

nuScenes Coordinate System

The nuScenes egovehicle coordinate frame is situated on the ground underneath the center of rear axle. Consider the pose of the LiDAR sensor in the egovehicle frame:

egovehicle_SE3_lidar.translation
array([0.94, 0.      , 1.84 ])

This means the LiDAR is on the center of the car, 1.84 meters above the ground. It is also almost one meter forward (+x) from the rear axle.

What about the relative orientation between the frames?

Rotation.from_matrix(egovehicle_SE3_lidar.rotation).as_euler('zyx', degrees=True)
array([-89.9,   1.4,   0.3])

We see that the LiDAR frame is basically rotated -90 degrees from the egovehicle frame

np.round(egovehicle_SE3_lidar.transform_point_cloud(np.eye(3)),2)
array([[ 0.95, -1.  ,  1.83],
       [ 1.94,  0.  ,  1.82],
       [ 0.97, -0.01,  2.84]])

Another Option for DCN: