Large-scale_Point_Cloud_Semantic_Segmentation

Progection Fusion

This project is a point cloud semantic segmentation network trained by the snapnet network on semantic3D data. The network first generates images for training on the point cloud and then uses other image semantic segmentation networks to train on the images. When processing new point cloud data, RGB images are automatically generated on the point cloud, and these images are labeled by the image semantic segmentation network. The network backprojects the results of the image segmentation to the point cloud.
The following GIF is the result of the network segmentation point cloud

SnapNet & Semantic 3D dataset

ShapNet

SnapNet's performance on the semantic 3D dataset is second only to SPGraph. It is the best network based on tensorflow in the field of point cloud segmentation. Then the snapnet network later introduced the robot version snapNet++ for target detection.

Semantic 3D dataset

They provide training and test data as a compressed ascii text file format of {x, y, z, intensity, r, g, b}. The basic facts are provided in the form of a single column ascii file, where the row id of the class labels corresponds to the point. 7zip is used for compression and is available for Windows and Linux.

Data download link：　http://www.semantic3d.net/view_dbase.php?chl=1

You need to put the downloaded point cloud data and tag data in the train folder.

Operating Environment

C++：　

Cython 0.29.1
PCL = 1.8
OpenMP
NanoFlann: nanoflann.hpp should be included in the include directory
Eigen: Eigen should also be included in the include directory

Python:

TensorFlow
TQDM, Scipy, Numpy ues the pip install

Building

cd pointcloud_tools
python setup.py install --home="."

Configuration file

{
    "train_input_dir":"path_to_directory_TRAIN",
    "test_input_dir":"path_to_directory_TEST",
    "train_results_root_dir":"where_to_put_training_products",
    "test_results_root_dir":"where_to_put_test_products",
    "images_dir":"images",

    training:true,

    "imsize":224,
    "voxel_size":0.1,

    "train_cam_number":10,
    "train_create_mesh" : true,
    "train_create_views" : true,
    "train_create_images" : true,

    "test_cam_number":10,
    "test_create_mesh" : true,
    "test_create_views" : true,
    "test_create_images" : true,

    "vgg_weight_init":"path_to_vgg_weights.npy",
    "batch_size" : 24,
    "learning_rate" : 1e-4,
    "epoch_nbr" : 100,
    "label_nbr" : 10,
    "input_ch" : 3,

    "train_rgb" : true,
    "train_composite" : true,
    "train_fusion" : true,

    "saver_directory_rgb" : "path_to_rgb_model_directory",
    "saver_directory_composite" : "path_to_composite_model_directory",
    "saver_directory_fusion" : "path_to_fusion_model_directory",
    "output_directory":"path_to_output_product_directory"
}

Processing training datas

The point cloud decimation
views and images generation
The images that will be generated when running the image generation script and the corresponding camera location files will be placed in './'folder. When the training script is run, the train_save file will be created containing the label point cloud and RGB point cloud, as well as the 2D image images generated on these point clouds, which will be used for training.

python3 sem3d_gen_images.py --config config.json

Train the models (rgb, composite and fusion) from scratch

Three files are generated during training to save the training results.'composite_save''fusion_save''rgb_save'

python3 sem3d_train_tf.py --config config.json

semantic on decimated clouds

The semantic predictions on images
back-projection on the decimated clouds
This script will generate tagged point cloud data（ply）

python3 sem3d_test_backproj.py --config config.json

Assign a Label to original point

generate the files at the Semantic 3D format
assign a label to each point of the original point cloud
This script marks the point of the corresponding row in the point cloud file as the Train ID.

python3 sem3d_test_to_sem3D_labels.py --config config.json

Using the network to inference on new data

Preprocessing point cloud data

First you need to use my las.py to convert your data into a model readable format.Modify the point cloud path in the script.

python las.py

Inference

Modify 'trainng model = false' in the Config.json file.
Add new data names to the sem3d_gen_images.py, sem3d_test_backproj.py, sem3d_test_to_sem3D_labels.py scripts
Run the script to process the point cloud.

python3 sem3d_gen_images.py --config config.json

Run the script to predict and output the ply file.

python3 sem3d_test_backproj.py --config config.json

Run the script to mark each point in the original point cloud.

python3 sem3d_test_to_sem3D_labels.py --config config.json

jtpils / large-scale_point_cloud_semantic_segmentation Goto Github PK