A package for semantic segmentation of 3D unorganized point clouds using Machine Learning techniques.
Semantic segmentation of point clouds is the process of classifying each point in a point cloud into different semantic classes, such as building, road, or vegetation. One approach to accomplish this is by using Machine Learning techniques such as Random Forest and Gradient Boosting classifiers.
Random Forest is a type of ensemble learning method that combines multiple decision trees to make predictions. In the context of semantic segmentation of point clouds, each decision tree in the Random Forest model would be trained to classify a point based on its features, such as its location and color. The final prediction for a point would be the majority vote of all the decision trees in the forest. Random Forest has been shown to be effective for semantic segmentation of point clouds due to its ability to handle high-dimensional and noisy data.
Gradient Boosting is another Machine Learning technique used for classification tasks. It gives a prediction model in the form of an ensemble of weak prediction models, which are typically decision trees. When a decision tree is the weak learner, the resulting algorithm is called gradient-boosted trees; it usually outperforms random forest. A gradient-boosted trees model is built in a stage-wise fashion as in other boosting methods, but it generalizes the other methods by allowing optimization of an arbitrary differentiable loss function.
You can install SemanticML from source by easily running:
git clone https://github.com/Yarroudh/SemanticML
cd SemanticML
python setup.py installAfter installation, you have a small program called sml. Use sml --help to see the detailed help:
Usage: sml [OPTIONS] COMMAND [ARGS]...
CLI tool to perform semantic segmentation of 3D point clouds using Machine
Learning techniques.
Options:
--help Show this message and exit.
Commands:
train Train the model for semantic segmentation of 3D point clouds.
predict Perform semantic segmentation using pre-trained model.
The process consists of two distinct steps or commands :
Model training is the process of using a set of labeled data, known as the training dataset, to adjust the parameters of Random Forest algorithm so that it can make accurate predictions on new, unseen data. The process of training a model involves providing the model with input-output pairs, where the input represents the features of the data and the output represents the desired label or prediction. The model then adjusts its internal parameters to minimize the difference between its predictions and the true labels. The goal is to find the set of parameters that result in the lowest prediction error on the training data.
This is done using the first command train. Use sml train --help to see the detailed help:
Usage: sml train [OPTIONS] CONFIG
Train the model for semantic segmentation of 3D point clouds.
Options:
--method [RandomForest|GradientBoosting]
Learning method for classification.
[default: RandomForest]
--help Show this message and exit.
The input data is a LAS file with specified features and classification field that respresents the label. The command takes one argument which is a JSON file that contains the features to use, the training data path and the algorithm parameters, as shown in this example:
{
"features": ["red", "blue", "green", "Verticality16", "Verticality8", "Linearity16", "Linearity8", "Planarity16", "Planarity8", "Surfacevariation5", "Numberneighbors10"],
"label": ["Ground", "Vegetation", "Rail", "Catenary pole", "Cable", "Infrastructure"],
"training_data": "C:/Users/Administrateur/Desktop/railway.las",
"parameters": {
"RandomForest": {
"n_estimators": [50],
"criterion": "entropy",
"max_depths": null,
"min_samples_split": 4,
"min_samples_leaf": 4,
"min_weight_fraction_leaf": 0,
"max_features": "sqrt",
"max_leaf_nodes": null,
"min_impurity_decrease": 0.0,
"bootstrap": true,
"oob_score": false,
"n_jobs": -1,
"random_state": null,
"verbose": 0,
"warm_start": false,
"class_weight": null,
"ccp_alpha": 0.0,
"max_samples": null
},
"GradientBoosting": {
"n_estimators": [50],
"loss":"log_loss",
"learning_rate":0.1,
"subsample": 1.0,
"criterion": "friedman_mse",
"min_samples_split": 2,
"min_samples_leaf": 1,
"min_weight_fraction_leaf": 0.0,
"max_depths": [3],
"min_impurity_decrease": 0.0,
"init": null,
"random_state": null,
"max_features": null,
"verbose": 0,
"max_leaf_nodes": null,
"warm_start": false,
"validation_fraction": 0.1,
"n_iter_no_change": null,
"tol": 1e-4,
"ccp_alpha": 0.0
}
}
}Thus, the command for Random Forest algorithm could be:
sml train --method RandomForest config.json
Or simply:
sml train config.json
You can also choose to use Gradient Boosting classifier by typing:
sml train --method GradientBoosting config.json
The output is the model with the best parameters, saved as a pickle file in ./output/model.
Pickling a model and saving it to disk allows you to save the state of the model, including all its trained parameters, so that it can be loaded and used again later without having to retrain the model from scratch. This is useful in cases where training a model takes a long time, or if you want to share a trained model with others.
Once the model is trained, it can be used to make predictions on new, unseen data. This is done by providing the model with input data, and the model generates an output, which is a LAS file with classification as a new scalar field.
This is done using the second command predict. Use sml predict --help to see the detailed help:
Usage: sml predict [OPTIONS] CONFIG POINTCLOUD MODEL
Perform semantic segmentation using pre-trained model.
Options:
--regularize BOOLEAN If checked the input data will be regularized.
[default: False]
-k INTEGER Number of neighbors to use if regularization is set.
[default: 10]
--filename PATH Write the classified point cloud in a .LAS file.
[required]
--help Show this message and exit.
sml predict config.json unclassified.las ./output/model/ne60_mdNone.pkl --filename classified.las
This uses the trained model stored as pickle file ne60_mdNone.pkl to perfoem semantic segmentation on unclassified.las. The output is named classified.las and can be found in the folder ./output/prediction.
The rendering of the Random Forest can be improved by an algorithm that reduces the noise. The principle of the algorithm is quite simple:
- For each point, we calculate the K-nearest neighbors
- We calculate the percentage of each class among these K-nearest neighbors
- If the majority class is present in the environment at more than X% percent, then the point is automatically assigned this class This allows to update the value of the classification for specific points to a value determined by a K-nearest neighbors vote.
sml predict config.json unclassified.las ./output/model/ne60_mdNone.pkl --filename classified.las --regularize True -k 30
In this example, regularization is enabled and the number of neighbors to use is 30.
- The training data should be classified with scalar field named
classification - The labeled and unseen data should contain all the features specified in the
configurationfile - Supported formats:
.LAS.LAZ
If you find our work useful in your research, please consider citing:
@inproceedings{NICE2020,
title={TESSERAE3D: A BENCHMARK FOR TESSERAE SEMANTIC SEGMENTATION IN 3D POINT CLOUDS},
author={Kharroubi, A., Van Wersch, L., Billen, R., and Poux, F.},
booktitle={ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., V-2-2021, 121–128},
year={2021}
}This software was developped by Kharroubi Abderrazzaq and Anass Yarroudh, researchers at the Geomatics Unit of the University of Liege. For more detailed information please contact us via [email protected] or [email protected], we are pleased to send you the necessary information.