Open3D-PointNet2-Semantic3D-master的运行

• 1.修改download_semantic3d.sh文件

#!/bin/bash

ans=`dpkg-query  -W p7zip-full`
if [ -z "$ans" ]; then
    echo "Please, install p7zip-full by running: sudo apt-get install p7zip-full"
    exit -1
fi

for i in `cat semantic3D_files.csv`
do
    output_file=`basename $i`
    echo Downloading ${output_file} ...
    #把wget $i改成：wget -c -N $i
    wget -c -N $i
    7z x ${output_file} -y
done

mv station1_xyz_intensity_rgb.txt neugasse_station1_xyz_intensity_rgb.txt

exit 0

• 2.修改preprocess.py文件

import os
import subprocess
import shutil
import open3d

from dataset.semantic_dataset import all_file_prefixes


def wc(file_name):
    out = subprocess.Popen(
        ["wc", "-l", file_name], stdout=subprocess.PIPE, stderr=subprocess.STDOUT
    ).communicate()[0]
    return int(out.partition(b" ")[0])


def prepend_line(file_name, line):
    with open(file_name, "r+") as f:
        content = f.read()
        f.seek(0, 0)
        f.write(line.rstrip("
") + "
" + content)


def point_cloud_txt_to_pcd(raw_dir, file_prefix):
    # File names
    txt_file = os.path.join(raw_dir, file_prefix + ".txt")
    pts_file = os.path.join(raw_dir, file_prefix + ".pts")
    pcd_file = os.path.join(raw_dir, file_prefix + ".pcd")

    # Skip if already done
    if os.path.isfile(pcd_file):
        print("pcd {} exists, skipped".format(pcd_file))
        return

    # .txt to .pts
    # We could just prepend the line count, however, there are some intensity value
    # which are non-integers.
    print("[txt->pts]")
    print("txt: {}".format(txt_file))
    print("pts: {}".format(pts_file))
    with open(txt_file, "r") as txt_f, open(pts_file, "w") as pts_f:
        for line in txt_f:
            # x, y, z, i, r, g, b
            tokens = line.split()
            tokens[3] = str(int(float(tokens[3])))
            line = " ".join(tokens)
            pts_f.write(line + "
")
    prepend_line(pts_file, str(wc(txt_file)))

    # .pts -> .pcd
    print("[pts->pcd]")
    print("pts: {}".format(pts_file))
    print("pcd: {}".format(pcd_file))
    """
    point_cloud = open3d.read_point_cloud(pts_file)
    open3d.write_point_cloud(pcd_file, point_cloud)
    改成：
    point_cloud = open3d.io.read_point_cloud(pts_file)
    open3d.io.write_point_cloud(pcd_file, point_cloud)
    """
    point_cloud = open3d.io.read_point_cloud(pts_file)
    open3d.io.write_point_cloud(pcd_file, point_cloud)
    os.remove(pts_file)


if __name__ == "__main__":
    # By default
    # raw data: "dataset/semantic_raw"
    current_dir = os.path.dirname(os.path.realpath(__file__))
    dataset_dir = os.path.join(current_dir, "dataset")
    raw_dir = os.path.join(dataset_dir, "semantic_raw")

    for file_prefix in all_file_prefixes:
        point_cloud_txt_to_pcd(raw_dir, file_prefix)

• 3. Convert txt to pcd file

Run

python preprocess.py

Open3D is able to read .pcd files much more efficiently.

• 4. Downsample

The downsampled dataset will be written to dataset/semantic_downsampled. Points with label 0 (unlabled) are excluded during downsampling.

downsample.py文件中的open3d.Vector3dVector()改为:

open3d.utility.Vector3dVector()

 5.open3d.voxel_down_sample_and_trace(
dense_pcd, voxel_size, min_bound, max_bound, False
)改成:

open3d.geometry.PointCloud.voxel_down_sample_and_trace(
        dense_pcd, voxel_size, min_bound, max_bound, False
    )

• 5. Compile TF Ops

cmake ..这一步遇到的错误：CMake Error at CMakeLists.txt:4 (cmake_minimum_required):

  CMake 3.8 or higher is required.  You are running version 3.5.1
解决办法：

升级cmake

参考：https://blog.csdn.net/weixin_43046653/article/details/86511157

问题还是没有得到解决。

直接复制pointnet++中编译好的.so文件到build directory.

(After compilation the following .so files shall be in the build directory.)

• 6. Train

Run

python train.py

• 7. Predict

Pick a checkpoint and run the predict.py script. The prediction dataset is configured by --set. Since PointNet2 only takes a few thousand points per forward pass, we need to sample from the prediction dataset multiple times to get a good coverage of the points. Each sample contains the few thousand points required by PointNet2. To specify the number of such samples per scene, use the --num_samples flag.

python predict.py --ckpt log/semantic/best_model_epoch_040.ckpt 
                  --set=validation 
                  --num_samples=500

The prediction results will be written to result/sparse.