代写代考 Perception for Autonomous Systems 31392:

Perception for Autonomous Systems 31392:
Image Feature Detection and Description
Lecturer: —PhD
10 Feb. 2020 DTU Electrical Engineering 2

Image Features
• Feature Detection:
Find the most “prominent” Points (areas) in an image.
The ones which are likely to be detected in other images, as well
• Feature Description:
Create a “unique” descriptor fingerprint for each Feature point
• Feature Matching:
Find correspondences among different images
10 Feb. 2020 DTU Electrical Engineering

Image Features
• A quick view
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Harris corner detector
• Corners are great for features
• Lines not so, why
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Harris corner detector • Harris can discriminate among edge, flat and corners
• Notice how by rotating it does not change • Linear algebra
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Harris corner detector
• So how are these eigenvalues useful?
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Harris corner detector
• So how are these eigenvalues useful?
Mikolajczyk, K., and Schmid, C., “A performance evaluation of local descriptors”,
IEEE Transactions on Pattern Analysis and Machine Intelligence, 10, 27, pp 1615–1630, 2005.
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Scale Space Theory
• So we can find corners.
• But how descriptive are these corners?
– Not really,
– Think that the roof of a building has corners – And you desk has corners…
• Finally a revelation:
Lindeberg, T. 1994. Scale-space theory: A basic tool for analysing structures at different scales. Journal of Applied Statistics, 21(2):224-270
Lindeberg, Tony (1998). “Feature detection with automatic scale selection”. International Journal of Computer Vision 30 (2): 79–116.
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Scale Space Theory
• We should be able to find these points which are prominent in different scales
• Create octaves with different scale among them
• Different blur level (Gaussian) in them
10 Feb. 2020 DTU Electrical Engineering

Feature Detection: Difference of Gaussians
• The main point is the difference of Gaussians • We can then do this on our scale-space:
10 Feb. 2020 DTU Electrical Engineering

The mother of Features – SIFT
• Ok, we’ve seen how we can find interest points, but how about matching?? • published the most influential paper in computer vision:
Lowe, D. Distinctive Image Features from Scale-Invariant Keypoints.International Journal of Computer Vision, 60, 2 , pp 91-110 (2004).
• How many people do you think have cited this? • 60000!!!!
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
Contains all:
• Detection
• Description
• Matching
Main Feature is that it is robust in: • Change of Translation
• Change in Scale
• Change in Rotation
• Change in 3D View Point • Change in Illumination
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
• Use DoG on Scale Space
– Only the maximum or minimum in a neighborhood are considered
– All octaves are investigated
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
• Use DoG on Scale Space
– Only the maximum or minimum in a neighborhood are considered
– All octaves are investigated
• Specifically,
– In groups of 3
– 2 Set of Points from each octave – 4 octaves -> 8 set of Points
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
Description
• Orientation:
– Based on the gradient
– Create a histogram of orientations Consisting of 36 bins (every 10 degrees)
• Fingerprint
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
Description
• Orientation • Fingerprint:
– Assume a 16 x 16 area around each Key Point
– Create histogram with 8 bins (as before)
– This gives out 128 values
• Note that the values are scaled for proximity
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
• Given some features from (let’s say) 2 Images:
• Lowe proposed the ALL-ALL Euclidean distance:
• For a 640×640 image we can get even 1000 features
• With 128 values each feature, you see how this can get ugly quickly
10 Feb. 2020 DTU Electrical Engineering

Scale Invariant Feature Transform – SIFT
• So our hero, proposed the usage of Kd-Trees: • Creation:
– You split the space in half based on distance
– Then again
– Then again,…. • Search:
– Start from top
– Is it closer to 1 than 2?
– Is it closer to 1.1 than 1.2?
– GO On until you find a feature
10 Feb. 2020 DTU Electrical Engineering

Feature points and areas
• A multitude of features have been proposed (specific for each application):
– SURF – BRISK – FREAK – MSER – ORB –…
10 Feb. 2020 DTU Electrical Engineering 18

Feature points and areas
• A multitude of features have been proposed (specific for each application):
– SURF – BRISK – FREAK – MSER – ORB –…
A special case to research
10 Feb. 2020 DTU Electrical Engineering 18

What applications do the Features have?
• More or less everything..
• Motion Estimation • Localization
• Photogrammetry • Image Retrieval
• Machine Learning – Object Detection – & Recognition
• Autonomous Driving
• More or less everything!
10 Feb. 2020 DTU Electrical Engineering

Image Feature Detection and Description
• What did we learn?
– Is that a good feature?
– How to get scale and rotational invariance in the features we get? – How to detect points of interest?
– How to describe the feature of points so,
– We can match them across multiple images.
– How can we use features?
10 Feb. 2020 DTU Electrical Engineering 20

Perception for Autonomous Systems 31392:
Image Feature Detection and Description
Lecturer: —PhD
10 Feb. 2020 DTU Electrical Engineering 21

程序代写 CS代考加微信: powcoder QQ: 1823890830 Email: powcoder@163.com

Related Posts