Computer Vision
Computer Vision: Replicating the human visual system in machine learning, granting them the ability of gaining understanding and reasoning from visual information images and videos
-Processing done mostly on image or videos
3D Computer Vision
One branch of computer vision
Traditional computer vision understands the world by referring to the fact that 3D space information is projected to the 2D image space, and from the visual information, performs an inference to the information needed
3D Computer vision, on the other hand can use the 3D data acquired through various sensors
How to acquire these 3D data?
(1)Stereo Images
Using stereo images, can acquire stereo depth
(2)Multi-view Images
Can acquire multi-view geometric using multi-view Imagess
(3)RGB-D camera or Point cloud camera
Using RGB-D camera or Lidar acquired Point cloud data can be acquired in real-time
Various 3D data can be acquired directly and processing such data, performs the task of understanding the situation and infer to the information needed
Understanding 3D
3D data acquired using various sensors or processes can be represented many ways by the computer
-Parametric (explicit) surface representation
Direct information about each point on the surface
The postitions of points, surface attributes(such as normal, colors, etc.) are directly stored
=> Point cloud is one of the explicit representation
-Implicit surface representation
3D surface, or 3D figure from the sensor is defined in the form of mathematical function representing the surface
In the function, the value that thresholds (or ‘cuts off’) is defined as the surface
-Implicit surfaces can represent more complex surface shapes compared to explicit surfaces and are particularly useful for smoothly changing surfaces
-Example:
Let’s say we want to represent a circumference of a circle
Explicit Representation: (x, y)
Enumerating information such as triange mesh, point.. and change the figure
Implicit Representation: Define it as a function, and the internal of a circle is F(x,y)<0 and define the external as F(x, y)>0 and the point that is 0 is the surface
-SDV is one of the ways to express this process
-Pointcloud data is a position in a 3rd-dimenson space, normal vector like figure information is explicitly defined, therefore has an explicit representation
Reference: https://www.youtube.com/watch?v=nSVOSIUYna4&list=PLubUquiqNQdN83-fPBzzViEEqohpdlwk2