Introduction to AFid
Advanced Feature IDentification (AFid)is a sophisticated software plugin (SDK)developed by Fluid Dynamic Sciences LLC. Originally conceived to enhance automated mesh generation by improving the application of mesh spacing rules, AFid leverages advanced machine vision technology, initially adapted from self-driving car systems, to offer robust 3D feature recognition across various industrial and medical sectors.
Unlike many contemporary solutions that rely on extensive training data and time-consuming neural networks, AFid is designed for efficiency and ease of integration. It requires only a single example of a feature to automatically recognise it within much larger and complex datasets, making it up to 720 times faster than manual processes in some applications. For instance, in a complex case study with ADL / Boeing 737, specifying CFD mesh refinement rules took an expert 4 hours, compared to only 20 seconds with AFid.
AFid is protected by US and worldwide patents, which covers the identification of features within geometry and the automation of refinement rules based on detected features. A patent continuation explicitly extends coverage from point clouds to neural network equivalents.
Technology: Leveraging Machine Vision without the Neural Network Overhead
AFid’s core strength lies in its use of non-neural network machine vision techniques to efficiently process 3D data. The technology operates on a generic data structure known as a point cloud.
A point cloud is the most fundamental representation of a shape, consisting of an unordered set of vertices (points) without explicit connectivity information. This representation is inherently generic and capable of accommodating digital data from a wide range of sources, including:
- Design Data: Digital CAD models.
- Scanned Data: In-service data from LIDAR, MRI, and CT scanners.
AFid’s patented approach to pattern matching is based on established, highly efficient computer vision algorithms:
- Voxelisation and Windowing: These techniques are used to pre-process large point clouds by partitioning the data into a grid of voxels, which significantly reduces the data size and enables efficient processing and spatial indexing.
- Fast Point Feature Histograms (FPFH): FPFH is employed to efficiently compute local feature descriptors. These descriptors capture the geometric properties (e.g., curvature, proximity) of a point and its neighbours, which are used to find correspondence between a reference feature and the larger scene point cloud.
- Random Sample Consensus (RANSAC): This iterative method is used to estimate the parameters of a mathematical model from a set of observed data that contains outliers. In AFid, RANSAC helps to robustly identify potential matches for a feature, even if the data is noisy or incomplete (e.g., in partial object detection).
- Iterative Closest Point (ICP): Once an initial match is found (often via RANSAC), ICP is used to precisely align the reference feature point cloud with the identified feature in the scene point cloud. The output of this process is a transformation matrix that defines the feature’s location and orientation.
The Advantage over Neural Network Approaches
Many modern feature recognition systems rely on deep learning and neural network (NN)architectures. While powerful, these approaches present significant drawbacks that AFid successfully mitigates:
| Feature | AFid (Non-NN) | Neural Network Approaches |
| Training Requirement | Only a single example is required (1-shot learning) | Extensive, labeled training datasets are required |
| Training Time | No extensive training time required | Significant time investment for training and validation |
| Accuracy | High, robust recognition invariant of transformation/orientation | Can be inaccurate if training data is insufficient or poorly labeled |
Computational Overhead | Runs quickly and efficiently on desktop computing equipment | Often requires powerful GPUs for training and inference |
AFid’s non-NN approach is fundamentally more efficient, particularly for industrial applications where time-to-solution is critical. By only requiring a single example, it removes the extensive effort and computational cost associated with collecting, labeling, and training large datasets—a process that can often lead to inaccurate or brittle models if not executed perfectly.
Applications and Integration
AFid’s point cloud-based, industry-agnostic architecture allows it to be integrated easily within existing software frameworks (e.g., via the AFid SDK which offers Python and C/C++integration) to provide feature recognition capability for numerous use cases:
- Computational Fluid Dynamics (CFD):Automatic application of complex mesh spacing rules on geometric features (e.g., vortex generators, wings, wheels), significantly speeding up pre-processing time.
- Quality Control & Manufacturing: Detecting non-conformities in printed circuit boards, checking the position or damage in medical implants (CT scan data), or comparing as-designed to as-commissioned data in industrial factories (LIDAR/CAD).
- Asset Management and Inventory: Finding and counting features like industrial rivets in structural models to automatically tag them, add them to a Bill of Materials, or remove them from the model.
- Incomplete Data Handling: The technology can detect partial objects from occluded or incomplete scan data (e.g., a portion of a vehicle), or match a 3D CAD model to a 3D point cloud reconstructed from a video.
The AFid SDK is designed for easy portability and is offered as a library packaged with full API documentation and examples. The demonstrator application allows users to test the functionality with their own point clouds, create, and save feature libraries. The SDK is available for partner evaluation from Q1 2025.
References
● US and Worldwide Patents: Fluid Dynamic Sciences LLC. Fast feature recognition and mesh generation in structural design, EP4573528A2.
● Website: Fluid Dynamic Sciences LLC. https://fluiddynamicsciences.com/