hotthoughts/cloudpose_pt

This is a pytorch implementation of "Learning Object Pose Estimation with Point Cloud"

CloudPose_pt

CloudPose pytorch version.

You can train&val with your own dataset.

Introduction

This repository contains the implementation of CloudPose in PyTorch.

From Paper: "Learning Object Pose Estimation with Point Cloud"

CloudPose is also available in Tensorflow

Requirements

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

Preparations

Data Version Control

All data and model files are version controlled, managed by DVC.

Make your own dataset

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...)

Dataset Structure

example:

train:

./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.

val:

./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.

Train a network

python train.py --batch_size 128 --log_dir log --num_point 1024 --num_class=18

Visualize loss

python -m tensorboard.main --logdir log --port=3111 --host=127.0.0.1

References

CloudPose

Code Structure from votenet

PointNet pytorch implementation