Machine learning
Learning from data to help us make better-informed decisions
How machine learning is going to help us?
A construction site is a complex non linear system where materials, people, logistics, plans and budgets. These inputs interact to coordinate the assembly of hundreds of thousands of elements in time and space. Such a complex non linear process operating at different levels of granularity and at different frequencies, is a result of a myriad of decisions that are not entirely captured by any single entity. In order to capture these latent relations in this high dimensional decision manifold, unique to the construction process, we utilize machine learning.
One of the ways the construction process can be modeled is using a BIM. A 4D BIM captures the ideal state evolution of a construction project. This when combined with onsite observations provided by reality capture, provides a representation that encodes both the geometric and semantic relations between the as-built vs as-designed. We utilize state-of-the-art machine learning models to encode these relationships and perform inference over them.
Machine learning research at Naska.AI
At Naska.AI we employ various flavors of machine learning that are adapted to the peculiarities of our domain and its corresponding data. We specialize in Geometric Deep Learning to build models that operate on data structures that can encode geometry, texture and semantics. We utilize representation learning methods to build abstractions of our domain specific data that can be utilized for various downstream learning tasks. Finally, we also focus on Active and Few shot learning methods to adapt the insights captured by our baseline models to new tasks.
The machinery of Graph Neural Networks allow us to capture and model BIM data, and combine this information with onsite observations represented by pointclouds. Since construction sites are extremely cluttered, the pointcloud data captured onsite is both partial and noisy. We analyze this data using advanced 3D machine learning models that can extract insights from partial and noisy construction 3D data. Despite utilizing these models, every construction project is unique, there by inhibiting the generalization capabilities of these models. We utilize active learning and few shot learning methods that allow us to adapt these models with limited labeled data thereby making the problem tractable and enhancing customer experience.
Graphical networks
Using graph structures to represent relationships between variables
in machine learning.
3D machine learning
Extracting meaningful insights while analyzing and modeling 3D data obtained from various sources.
Active learning
Selective data labeling to improve machine learning performance with limited labeled data.
Ongoing projects
BIM-Graph Neural Network
Comparing elements in as-designed Building Information Models (ad-BIMs) with their registered 3D point clouds facilitates automatic construction quality assessments. However, independent analysis of elements can lead to incomplete assessments due to noise in on-site data or occlusions. To address this, we introduce BIM-Graph Neural Network (BIM-GNN) – a novel approach to element-wise quality assessment that leverages the semantics in ad-BIM to:
1. Enhance the performance of machine learning models.
2. Be robust to data incompleteness due to noise, occlusion, and poor scanning.
3. Infer the status of partially observed elements.
4. Infer the status of unobserved elements associated semantically.
5. Reduce manual labeling and post-processing efforts.
BIM-GNN allows us to say more with less.
