Performance Evaluation of Different Devices and Algorithms for Modelling Small Artefact

3D reconstruction and modelling play important roles in various applications, specifically in heritage preservation. With the aid of suitable hardware like the 3D sensors as well as respective data processing methods, the work has become more feasible in realizing the aim to conserve and preserve more artefacts. However, too many choices and alternatives might lead to different results which may affecting the preservation purpose. The objective of this work is to analyze and evaluate the performance of different devices and algorithms for small artefact modelling. Two 3D sensors, iPhone 13 Pro Max LiDAR and Structure sensor were selected to collect data of small artefact to be reconstructed and modelled. Two main, important surface reconstruction algorithms which are Poisson and Ball-Pivoting methods were also selected to be tested. Specifications of the sensors’ capabilities as well as modelling results of the artefact are examined. Different parameters of the algorithms were selected to study their effect. These findings will help to learn more about 3D sensors and the suitable modelling methods in making them better for usage in a variety of areas, including archaeology, architecture, and the protection of cultural heritage.


Introduction
The use of 3D sensors for historical object preservation has become increasingly popular in recent years.3D sensors can capture detailed 3D models of historical artefacts and sites, providing valuable information for researchers, historians, and younger generations.However, not all 3D sensors are created equal, and it is important to evaluate the performance of different systems to determine which ones are best suited for artefact preservation.
There are different types of 3D sensors that can be used for this purpose, each with its own advantages and limitations.Time-of-Flight (ToF) sensors use laser pulses to measure the time for the light to bounce back to the sensor [1].It can produce accurate 3D point clouds and can operate in a wide range of lighting conditions, making it suitable for both indoor and outdoor environments [2].The light detection and ranging (LiDAR) sensor available on the iPhone 13 Pro Max is an example of a ToF sensor which can accurately measure the distance to objects of interest in the environment and create a detailed 3D data.It has a range of up to 5 meters and can scan at a rate of up to 30 frames per second [3].
On the other hand, another 3D sensor that is available is the Structure sensor.It uses a combination of an RGB camera, infrared sensor, and structured light technology to scan and create 3D data.It has a range of about 4 meters and can scan at a rate of up to 30 frames per second as well [4].Structure sensor has been preferred for applications in medical field as well as human lifestyle [5].
Often, these 3D data generated by the above-mentioned sensors are dense in information and thus, suitable methodology needs to be applied in processing and modelling them.Due to the rapid development in this area of application, various methods have been developed to assist in 3D point cloud data processing.Thus, the best approach with the suitable sensor combination specifically in modelling small artefact is studied here.

Methodology
In this research, different 3D sensors and algorithms were tested to evaluate their performance.For the device, two 3D sensors were selected; iPhone 13 Pro Max LiDAR and Structure sensor.Both sensors were chosen and preferred specifically in this application of small artefact modelling due to their portability and ability to produce very high point cloud density.Figure 1  Meanwhile, for the reconstruction and modelling method from point clouds, Ball-Pivoting Algorithm (BPA) and Poisson algorithm are the two common techniques used.Thus, these algorithms were chosen to be further analyzed.The BPA, which iteratively generates a surface mesh by pivoting balls over the input points, was first proposed in [7].However, several factors that can lead to inaccurate surface reconstruction which are low density and manifold's curvature as shown in Figure 2.

Figure 2.
The Ball-Pivoting Algorithm (BPA) [7] On the other hand, the Poisson algorithm, described in [8], formulates surface reconstruction as a Poisson equation and solves it using a variational approach.By integrating the gradient of a scalar In order to evaluate the selected devices and algorithms performance, several processes were conducted, which includes data acquisition, data processing, data modelling and reconstruction as well as data analysis and evaluation.Figure 4 shows the block diagram highlighting the general processes involved in this study.

Figure 4. Overall processes
Due to the specific application in this work for historical object preservation, an artefact is chosen to be scanned for data acquisition.A historical tombstone located in Muzium Kota Kayang, Perlis, is selected.Figure 5 shows the image of the tombstone.To standardize the data acquisition process, a base and a ruler are used as the reference during collecting and scanning the data.

Figure 5. The selected tombstone
For data processing, CloudCompare and MATLAB software are used.Some pre-processing methods were done in CloudCompare before MATLAB is used for reconstruction and modelling.Figure 6 shows the raw data collected from iPhone LiDAR and Structure sensor.Both devices provide good 3D point cloud results but there is a difference in the data, where the iPhone LiDAR produces data that is integrated with the true colour information of the object while the Structure sensor only provides a 3D point cloud without the colour information.To ensure that the variables and parameters are constant, both point cloud data have been cropped to a smaller one as shown as in Figure 7, concentrating on the tombstone itself.This will guarantee that the number of point clouds are on the same scanning size.Results of data reconstruction and modelling as well as the evaluation are analysed in the next section.

Results and Discussion
Poisson and BPA were applied to the pre-processed point cloud data to reconstruct and model the artefact.By adjusting the control parameters impacting the reconstruction quality, their performances are examined.For Poisson, the process of reconstructing the surface can be significantly impacted by the octree's value.The point cloud data is organized and efficiently represented using the octree data structure.Based on the quantity and distribution of the points, it subdivides the larger octants into smaller ones.The amount of detail in the rebuilt surface is influenced by the resolution of the octree, which is defined by the size of the octants.More complicated surface features can be captured and recreated in greater detail with a finer resolution and smaller octants.Figure 8 shows the reconstruction with default octree level of 8 implemented to both 3D point cloud data, while Figure 9 shows the reconstruction results with octree 5, 7 and 10 level.From here, it can be seen that Structure sensor produces more smooth surfaces but lack of detailing.Another important parameter that can be analyzed is the gridstep, which describes the size or resolution of a regular grid that is superimposed on the 3D environment.The grid is created by uniformly dividing the area into squares with fixed step sizes along each axis.The rebuilt mesh's level of detail is influenced by the gridstep, which controls the distance between grid points.A higher resolution grid is produced by smaller gridstep values as shown in Figure 10, allowing for a more accurate surface representation.For BPA, selecting the correct radius is essential.The radius establishes the greatest separation that can be maintained throughout the ball-pivoting operation between a pivot point and its surrounding points.The reconstruction results for various radius value can be seen in Figure 11, where it can be summarized that the smaller the radius, the more detailed the surface reconstruction.Overall, the reconstruction results of iPhone 13 Pro Max LiDAR data has rough surfaces but they produce a more meshing detail, while Structure sensor data produces more smoother modelling.This may be because of the higher density of point cloud generated by the device.Thus, the best combination of sensor and algorithm for modelling will depends on the higher level of details of the artefact as well.

Conclusion and Future Work
In conclusion, both iPhone 13 Pro Max LiDAR and Structure sensor have proven to be adequate in scanning 3D data representing an artefact.The selected Poisson and Ball-Pivoting surface reconstruction methods can be used in the reconstruction process to model the historical artefact.In the future, fusion and improvement of these algorithms can be performed.From here, a more precise and aesthetically pleasing model can be produced where approaches to combine their advantages in maximizing the performance can be performed.Furthermore, more testing can be done to other artefacts, perhaps in modelling a more complex, imperfect historical objects, in concluding the best device and algorithm.

Figure 1 .
shows the iPhone 13 Pro Max LiDAR and Structure sensor selected in this work.(a) Structure sensor and its attachment with an iPad[5]; (b) iPhone 13 Pro Max LiDAR[6]

Figure 6 .
The raw data collected from: (a) iPhone LiDAR; (b) Structure sensor

Figure 11 .
Reconstruction results at BPA radius of: (a) 0.02 m, (b) 0.01 m and (c) 0.007 m for iPhone LiDAR data (left) and Structure sensor data (right)