Scientific Paper: An Algorithm For Ordered Triangular Mesh Regularization
Authors: Claudio Tortorici, Naoufel Werghiy, Stefano Berretti
Over the years, computer scientists have developed several ways of representing data collected from the 3D world to generate virtual reality, improve object recognition, and enable self-driving cars. Various approaches suffered from either generating too much data or struggled to represent minute details accurately.
Now researchers at the UAE's Technology Innovation Institute have developed an innovative approach they call CSIOR that efficiently represents the subtle details of 3D textures – such as the weave of the fabric, the fine distinctions in ancient pottery from different cultures, or the 3D shape of a face.
"It's like if you found a way to create a higher quality image from raw camera data, you could get better results from other algorithms that analyzed the picture," said TII senior researcher Claudio Tortorici.
This could improve 3D reality capture and improve virtual reality applications in the short run. Eventually, this could lead to improvements in AI algorithms used for understanding the world around us.
Representing the world
Drone aircraft commonly use LiDAR or stereoscopic cameras to capture point cloud data about the world. A point cloud represents the distances between the drone and points on the surfaces of objects in the vicinity. But this does not explicitly capture the relationship between points. Two points near each other might be part of the same objects or parts of two separate objects right next to each other.
So, computer scientists have developed several ways of explicitly capturing this information – voxels and mesh manifolds. A voxel represents the world using 3D volume elements. This is the 3D equivalent of a pixel and bitmapped graphics. This is an excellent way of representing precise 3D structural data for domains like engineering. However, voxels also waste data resources for representing empty space or the depth of objects that you cannot see or even capture with LiDAR.
A voxel data set also grows in 3D dimensions, which means that capturing an object at twice the scale increases the data by 8-times. This pushes the need for memory, processing power, and networking requirements.
Another common technique is to transform raw point cloud data into a mesh manifold that uses triangles to connect the dots of the point cloud. A mesh manifold is like the 3D equivalent of vector graphics. This represents the data far more efficiently.
Mesh manifolds have been around for years. However, they often used triangles of varied sizes. This leads to two limitations. First, they lose some of the finer details in representing the textures of objects. Second, these irregularly shaped triangles confound AI and analytics algorithms. "It would be like trying to analyze a picture composed of different sized pixels," said Tortorici. "We are trying to regularize the size of the triangles."
A better mesh
The new technique is called the circle-surface intersection ordered resampling (CSIOR) algorithm because it uses a circular process for selecting the next bit to process. It applies an iterative process that enlarges the surface's coverage expanding out in a circle from a seed point.
The process generate a mesh manifold composed of triangles of the same shape and size. This brings uniformity to the data set. It could also simplify many types of graphics processing applications, such as cutting the outline of a face out of a 3D data set so it could be analyzed or pasted into another 3D scene.
Other mesh manifold all suffer from one of two problems. In some cases, they lose the intricate details on the surface. In others, they are good at keeping information but with different-sized triangles. "That would be like using irregular-sized pixel," Tortorici said.
CSIOR generates data sets that are the same size as other mesh manifold processing techniques. But they are much better at representing far more subtle textural details in objects. For example, a cup from the Chinese empire would have a completely different corrugation than one from the Roman empire.
Down the road, Tortorici believes that this technique could improve the efficiency of other algorithms used for filtering, convolutional processing, machine learning, and AI. The CSIOR algorithm does not do these directly but helps other algorithms get better results on these tasks
"A regularized mesh could allow us to run algorithms that could not run on irregular meshes," Tortorici said. "We believe this will improve results for algorithms that retrieve the pattern of a surface, classify surfaces, and segment surfaces."
We should include this graphic: