Search

Labelled point clouds are crucial to train supervised Deep Learning (DL) methods used for semantic segmentation. The objective of this research is to quantify discordances between the labels made by different people in order to assess whether such discordances can influence the success rates of a DL based semantic segmentation algorithm. An urban point cloud of 30 m road length in Santiago de Compostela (Spain) was labelled two times by ten persons. Discordances and its significance in manual labelling between individuals and rounds were calculated. In addition, a ratio test to signify discordance and concordance was proposed. Results show that most of the points were labelled accordingly with the same class by all the people. However, there were many points that were labelled with two or more classes. Class curb presented 5.9% of discordant points and 3.2 discordances for each point with concordance by all people. In addition, the percentage of significative labelling differences of the class curb was 86.7% comparing all the people in the same round and 100% comparing rounds of each person. Analysing the semantic segmentation results with a DL based algorithm, PointNet++, the percentage of concordance points are related with F-score value in R2 = 0.765, posing that manual labelling has significant impact on results of DL-based semantic segmentation methods. Xunta de Galicia | Ref. ED481B-2019-061 Ministerio de Ciencia e Innovación | Ref. PID2019-105221RB-C43 Ministère de l’Economie of the G. D. of Luxembourg | Ref. SOLSTICE 2019-05-030-24 Universidade de Vigo/CISUG

This study investigates the inability of two popular data splitting techniques: train/test split and k-fold cross-validation that are to create training and validation data sets, and to achieve sufficient generality for supervised deep learning (DL) methods. This failure is mainly caused by their limited ability of new data creation. In response, the bootstrap is a computer based statistical resampling method that has been used efficiently for estimating the distribution of a sample estimator and to assess a model without having knowledge about the population. This paper couples cross-validation and bootstrap to have their respective advantages in view of data generation strategy and to achieve better generalization of a DL model. This paper contributes by: (i) developing an algorithm for better selection of training and validation data sets, (ii) exploring the potential of bootstrap for drawing statistical inference on the necessary performance metrics (e.g., mean square error), and (iii) introducing a method that can assess and improve the efficiency of a DL model. The proposed method is applied for semantic segmentation and is demonstrated via a DL based classification algorithm, PointNet, through aerial laser scanning point cloud data.

A validation data set plays a pivotal role in tweaking a machine learning model trained in a supervised manner. Many existing algorithms select a part of available data by using random sampling to produce a validation set. However, this approach can be prone to overfitting. One should follow careful data splitting to have reliable training and validation sets that can produce a generalized model with a good performance for the unseen (test) data. Data splitting based on resampling techniques involves repeatedly drawing samples from the available data. Hence, resampling methods can give better generalization power to a model, because they can produce and use many training and/or validation sets. These techniques are computationally expensive, but with increasingly available high-performance computing facilities, one can exploit them. Though a multitude of resampling methods exist, investigation of their influence on the generality of deep learning (DL) algorithms is limited due to its non-linear black-box nature. This paper contributes by: (1) investigating the generalization capability of the four most popular resampling methods: k-fold cross-validation (k-CV), repeated k-CV (Rk-CV), Monte Carlo CV (MC-CV) and bootstrap for creating training and validation data sets used for developing, training and validating DL based point cloud classifiers (e.g., PointNet; Qi et al., 2017a), (2) justifying Mean Square Error (MSE) as a statistically consistent estimator, and (3) exploring the use of MSE as a reliable performance metric for supervised DL. Experiments in this paper are performed on both synthetic and real-world aerial laser scanning (ALS) point clouds.

Precise ground surface topography is crucial for 3D city analysis, digital terrain modeling, natural disaster monitoring, high-density map generation, and autonomous navigation to name a few. Deep learning (DL; LeCun, et al., 2015), a division of machine learning (ML), has been achieving unparalleled success in image processing, and recently demonstrated a huge potential for point cloud analysis. This article presents a feature-based DL algorithm that classifies ground and non-ground points in aerial laser scanning point clouds. Recent advancements of remote sensing technologies make it possible digitizing the real world in a near automated fashion. LiDAR (Light Detection and Ranging) based point clouds that are a type of remotely sensed georeferenced data, providing detailed 3D information on objects and environment have been recognized as one of the most powerful means of digitization. Unlike imagery, point clouds are unstructured, sparse and of irregular data format which creates many challenges, but also provides huge opportunities for capturing geometric details of scanned surfaces with millimeter accuracy. Classifying and separating non-ground points from ground points largely reduce data volumes for consecutive analyses of either ground or non-ground surfaces, which consequently saves cost and labor, and simplifies further analysis.

Most deep learning (DL) methods that are not end-to-end use several multi-scale and multi-type hand-crafted features that make the network challenging, more computationally intensive and vulnerable to overfitting. Furthermore, reliance on empirically-based feature dimensionality reduction may lead to misclassification. In contrast, efficient feature management can reduce storage and computational complexities, builds better classifiers, and improves overall performance. Principal Component Analysis (PCA) is a well-known dimension reduction technique that has been used for feature extraction. This paper presents a two-step PCA based feature extraction algorithm that employs a variant of feature-based PointNet (Qi et al., 2017a) for point cloud classification. This paper extends the PointNet framework for use on large-scale aerial LiDAR data, and contributes by (i) developing a new feature extraction algorithm, (ii) exploring the impact of dimensionality reduction in feature extraction, and (iii) introducing a non-end-to-end PointNet variant for per point classification in point clouds. This is demonstrated on aerial laser scanning (ALS) point clouds. The algorithm successfully reduces the dimension of the feature space without sacrificing performance, as benchmarked against the original PointNet algorithm. When tested on the well-known Vaihingen data set, the proposed algorithm achieves an Overall Accuracy (OA) of 74.64% by using 9 input vectors and 14 shape features, whereas with the same 9 input vectors and only 5PCs (principal components built by the 14 shape features) it actually achieves a higher OA of 75.36% which demonstrates the effect of efficient dimensionality reduction.