CloudPose pytorch version.
You can train&val with your own dataset.
This repository contains the implementation of CloudPose in PyTorch.
From Paper: "Learning Object Pose Estimation with Point Cloud"
CloudPose is also available in Tensorflow
Pytorch(tested with 1.1.0--1.5.0) on CUDA(tested with 10.0--10.2)
Tensorflow (tested with 1.14.0) for TensorBoard
opencv-python
All data and model files are version controlled, managed by DVC.
0000_seg.npy, 0001_seg.npy...respectively are two-dimensional numpy arrays containing 1024*channel, representing the points in the scene point cloud that have been segmented and downsampled to 1024.
0000_pos.npy, 0001_pos.npy ... are one-dimensional numpy arrays containing 17 elements, where [0:12] represents the first three rows and the first four columns of the rigid body transformation matrix, and the last element is the point cloud block Category (0, 1, 2...)
example:
./mydataset/my_train/0000_seg.npy
./mydataset/my_train/0000_pos.npy
./mydataset/my_train/0001_seg.npy
./mydataset/my_train/0001_pos.npy
etc.
./mydataset/my_val/8000_seg.npy
./mydataset/my_val/8000_pos.npy
./mydataset/my_val/8001_seg.npy
./mydataset/my_val/8001_pos.npy
etc.
python train.py --batch_size 128 --log_dir log --num_point 1024 --num_class=18
python -m tensorboard.main --logdir log --port=3111 --host=127.0.0.1
Code Structure from votenet
PointNet pytorch implementation