Q&A - Astraea-Software Co. LTD


Here are some of the questions we have received regarding our AI technology that enables 3D shape recognition instead of 2D.

About input data

What is the format of the input data for 3D shape recognition?
→The input data is 3D CAD data in STL format or OBJ format. (triangular surface shape data).

Is it possible to recognize 3D shapes from data created by scanners?
→If you have a scanner that can create OBJ format or STL format data files with a certain level of accuracy (number of nodes, number of coordinates), you can create the appropriate input data to enable 3D shape recognition.

In order to recognize 3D shapes, how many vertices (coordinate points) of CAD data are required?
→For example, a bolt model requires 1,000 to 2,000 vertices, and a car body model requires about 6,000 vertices. However, this is just a guideline and will be adjusted to the appropriate number according to your purpose.

In the demo site, 3D shape recognition is performed by selecting a file. Can I upload my own files for 3D shape recognition?
→Yes, it is possible. We do not allow user-uploaded files because the input should be preprocessed and should satisfy some requirements so that the AI can work properly. If you want to use your own data for 3D shape recognition, please contact us.


How to build a 3D AI model

What are the differences between 2D image recognition and 3D shape recognition methods?
→Image recognition is a technology in which an AI model recognizes a 2D image, while 3D shape recognition recognizes the 3D shape itself. Therefore, 3D shape recognition does not use image data, but uses only 3D shape data output from 3D CAD and other sources to train the AI model.
For details, please refer to the Research page of our website.

Is it possible to recognize the size (dimensions) of 3D data?
→ It is possible to recognize the size of a shape based on the coordinate value data of the input 3D data and search for a shape data with a similar size.
However, depending on the output CAD or scanner, the units of the coordinate values in the shape file (such as millimeters) may not match the actual units, so you may need to adjust the unit conversion.

・Is it possible to recognize the color of 3D data?
→ The AI model on the demo site does not recognize colors, but we can use the color code (RGB) as input features an let the AI model learn from them.


About Training an AI models

・How many models are required for training (for a classification AI)?
→ If you have about 20 patterns of 3D shape data for one classification class, you can train an AI model using data augmentation (a method to increase the number of patterns by rotating, moving, or flipping the 3D shape data).In general, the more data you have, the more accurate the classification becomes.

How long does it take for training?
→For example, for bolt shape recognition, it takes 2 to 3 hours to train 10,000 epochs with one training data.
In addition, since this process is repeated to improve the accuracy, we estimate that it takes about one week to complete the training.
The time listed above was measured using the computer with the following specifications.
CPU: Xeon Gold 6234 (8Core, 3.3GHz)
Memory: 192GB
GPU: NVIDIA QuadroRTX8000 (48GB)


Evaluation of the model

・How can I use the trained models?
→ AI functions can be provided via WebAPI. The AI functions can be provided via WebAPI, which allows for flexible use such as embedding them in customer applications or calling them from batch programs.

What is the accuracy level of 3D shape recognition? Can it be done in millimeters?
→ It depends on the size of the target model and the area of interest, so it is difficult to say in general.
Please contact us through the inquiry form and we will respond to your request.


Enhancements to the model

・Is it possible to increase the number of shapes (classification classes) to be recognized for 3D shape recognition?
→If you want to increase the number of classification classes for shape recognition, you will need to re-train the AI model.
However, when adding objects of different sizes, retraining is not necessary.


The above Q & A contains only the main questions and answers.
If you have any questions that are not mentioned, we would appreciate it if you could contact us.