Resolvent

  

 

 

Would you like a 3D-model mapping the geometrical (ie size) variation of your product, directly in your CAD system? Use it to find the statistical variance or correlation between on any quantity with a few mouse-clicks! Often a limited amount of point cloud scans can represent your product in great detail, read on to see how.


Customized tools based on 3D point clouds and machine learning
 

3D clouds and machine learningIt requires the right combination of vision, data and skills to make the design and production benefit from a large amount of data, which point clouds typically are.

 

Doing it right, you can benefit in several areas:

  • Optimize your product design, based on actual variations in manufactured products
  • …or even actual variation in the environment that your product should live in, ie the hands of persons
  • Segmented production related to actual shape and dimensions of output

 

At resolvent we enable our customers to utilize their large data sets, i.e. 3D point clouds obtained via scanning in ie quality control, by computing statistics, providing tools for data classification and establishing correspondence on whatever geometrical measure in an App integrated into their favorite CAD system. In the following we dig a bit deeper into how we approach it.

 

3D point clouds and pre-processing

Statistical analysis of scanned three dimensional shapes may require a lot manual work if the data has not been properly pre-processed. For many companies this becomes the limiting factor that hinters them in extracting valuable information from large data sets useful for improving the quality of their products.

 

3D

Automation (establish correspondence)

For simplicity consider analyzing the statistics of hands. E.g. measuring the length of the index finger, which would consist of manually tagging its tip and a point at its base, computing the distance between them and repeating this for all samples before finally calculating statistics. The key to improving this process is to initially establish correspondence between all meshes in the data set.

 

 

Figure 1: Triangulated point-mesh from a hand scan with indication of index finger length as measured from the data. 

 

Optimization of correspondence

Establishing correspondence between a reference mesh and target surface consists of morphing the reference mesh-points onto the target surface by assuming a parametric deformation model and then finding the parameters, which results in the best fit. This is done by solving an optimization problem. It does not have a unique solution and would result in a low-quality correspondence and may lead to erroneous statistics. 

                                                                                       

 

3D correspondenceFigure 2: Correspondence is established by morphing a reference mesh onto all compatible samples. Here exemplified by morphing a reference (grey) onto single target (green). Before morphing (left) and after morphing (right). 

 

Deformation fields-Landmarks 

A major improvement is achieved by incorporating a few known deformations at the most significant features of the shapes (landmarks) into the optimization as constrains. E.g. we could enforce that points at fingertips should map to points at fingertips. Fortunately, we only need to do this with a few samples, from which we can then construct a model which automatically morphs with perfect landmark correspondence. Using this deformation-model it is possible to automatically establish a high-quality correspondence for the processing of the larger batch of shapes.


 

Classification via Machine Learning models

In many cases it is also necessary to subdivide the data set into categories before attempting to establish correspondence. E.g. in the unlikely event that we would try to morph a mesh of a 5-finger hand to a 6-finger hand, we cannot hope to achieve a very meaning full correspondence. To resolve this, we beforehand need to subdivide the data set into compatible sets; 5-finger hands, 6-finger hands etc. Such classification is efficiently carried out using state-of-the machine learning models based on convolution neural networks while preserving the sparsity of data by representing the point clouds as octrees.

 

Take aways

The above workflow allows advanced statistics to be extracted directly from the shape in an intuitive manner, a few examples come to mind when keeping with the hand example:

  • Find the regions where the hands deviate the most, and incorporate that into your design of gloves
  • Easily segmented hand dimensions into groups, and segment your design based on this
  • Visualize variations directly on the hand geometry, to better understand and communicate this in your team and to clients

 

 

At resolvent we combine the industry-specific knowledge of our customers with the broad experience of our team, and implement it in state of the art Multiphysical Simulation and Data Analysis tools.