3D Vision Flashcards
3D Vision:
Name a difficulty when real world scenarios are projected from 3D to 2D in camera images.
Depth estimation
What are the main two types for 3D Data Acquisition?
Passive and Active range sensing
How many types of passive methods are there?
4
Shape from:
-stereo;
-motion;
shading;
-focus.
What are the pros of shape from stereo?
- Cheap (use cameras)
- Fast acquisition
What are the cons of shape from stereo?
- Highly dependant on
correspondences quality - Still challenging
What are the characteristics of Shape from motion?
– Similar to stereovision in many ways
– Successive images might be considered as stereo pairs
– With texture, possible to find correspondences (matching techniques, optical flow…) and find fundamental and
essential matrix.
What are the characteristics of shape from shading?
- given a continuous surface, and known illumination: intensity variation in the surfaces depends on its orientation
- Since most surfaces are not uniform and lighting is difficult to control, normally it is combined with other methods
What is shape from focus?
- objects away from focal point are out of focus;
- with different images with different focus its possible to extract depth information
How many types of Active range sensing are there?
- Structured Light Systems
- Laser Range Finder – Time of Flight
How does Structured Light Techniques are implemented?
-Projection of a known pattern
- Acquisition with camera, 3D from pattern deformation in scene.
What are the only positive point about Structured Light Techniques?
It is very accurate
What are the cons of Structured Light Techniques?
- Takes time (often need to scan through an area)
-Sensitive to environment brightness, usually only
implemented in dark or indoor areas. - Short range
When are the Laser Range Finders (LRF) used for?
They are used for larger areas (buildings, rooms)
How do Laser Range Finders (LRF) work?
Working principle:
light pulse time of flight - laser ray in -> reflect on object -> laser ray out
- phase shift: amplitude of frequency modulation - comparison of phases
What are the LRF pros?
- independent from external lighting;
- no need of texture in scene;
-provide directly 3D measurements
What are the LRF cons?
They are:
-expensive;
-large sensors = aquisition more difficult;
-limited spatial resolution;
-no color texture map
How do 3D ToF Cameras work?
They phase shift principle of emitted and received infrared light to measure depth
How could we define the perfomance and cost of Structured Light:
It has:
- best depth accuracy | shortest range | require dark environment
- Highest cost
How could we define the perfomance and cost of Time of Flight (ToF):
-performance is up to hundred meters. depending on emitting
power
- moderate cost
How could we define the perfomance and cost Camera Array:
- largest depth error | range depend on baseline (dist. between cameras = usually around 10 m)| require bright environment
- lowest cost | Development mainly on software side
In a nutshell Which perform better in the following categories, Active (Range - TOF) or Passive (camera arrays)?
Cost, Acquisition, Depth error, Texture map, Lighting, Texture relevance and 3D processing
Cost, Acquisition, Depth error, Texture map (first 4 categories):
Intensity (camera arrays) perform better - Passive
Lighting, Texture Relevance and 3D Processing (last 3 categories):
Are better performed by Range (TOF) -Active
What type does Kinect use?
Active - infrared pattern.
What composes a Kinect?
The kinect has:
- a multi-array mic
- 3D Depth sensors
- RGB camera
- motorized tilt
Where can 3D Vision be applied?
Robotics:
-Navigation, localization, mapping, avoiding collision
AR / VR:
- sensing real 3D environments and reconstructing them in the virtual world
To what the AR / VR Interaction’s devices must respond accurately?
To the 3D movement, therefore needing high-performance depth sensor
What is a Range image?
-It is a rectangular array of numbers that quantifies the distance from the sensor to the surfaces within the field of view.
-It isAlso referred as depth image and easily
transform to cloud of points.
How do Ranges differ from intensity to range images?
-In intensity images: edges related to intensity changes (due to
geometry or aspect - for example colour or shadow)
-In range images: jump or step edge; roof or crease edge; smooth edge
What is Registration?
It estimates Rigid Body Transform that minimize the distance
between 2 scans
What is the ICP (Iterative Closest Point) and how does it work?
It is an algortithm used to perform the registration:
– Find closest point
– Compute transform that minimizes error
– Repeat until ending condition.
What are the ICP problems?
- algorithm may fall in local minima
- typically requires an initial guess
- may result in many outliers
Besides ICP, it can also be done registration from clouds to surfaces: How many of these are there?
– Non parametric curves (triangles,…)
– Parametric curves (cylinders, quadrics, …)
What are some examples of triangulation algorithms?
Delaunay triangulation 2D;
Marching cubes;
Marching triangles;
Ball-pivoting;
Poisson Surface Reconstruction;
Moving least-squares (MLS)
How does the Delaunay triangulation work?
For a set of 2D points P ensure that none points of the set
is inside the circumcircle of any triangle.
What is Zippering?
- remove overlapping portion of meshes
- clip mesh together
- remove triangles
What is Texture Mapping?
Some 3D reconstruction techniques that provide automatically texture like for example
– Shape from …
– Structured Light Techniques
Do all kinds of reconstruction techniques provide texture?
No.
For example: initial Laser Range Finder does not provide texture
