SIFT FLOW: Dense Correspondence across Scenes and

Transcription

SIFT FLOW: Dense Correspondence across Scenes and
FACE DETECTION
Lab on Project Based Learning
May 2011
Xin Huang, Mark Ison, Daniel Martínez
Visual Perception
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Contents
• Introduction
• Exploration into local invariant features
• The final method
• Results
• Conclusions
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Contents
• Introduction
• Exploration into local invariant features
• The final method
• Results
• Conclusions
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Introduction
• Karlos Arguiñano’s TV show.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Introduction
Which is our
goal?
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Introduction
• Which is our goal? -> Detect presence of a
face.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Introduction
Problem
definition
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Introduction
• Problem definition:
• How can we use local invariant feature
descriptors to detect Arguinano’s face.
• Use MATLAB programming language.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Introduction
• Constraints of the project:
• Use local invariant features descriptors.
• Use MATLAB programming language.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Contents
• Introduction
• Exploration into local invariant features
• The final method
• Results
• Conclusions
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
First approaches:
1.
2.
3.
4.
5.
SURF matching.
Template database SURF matching.
Clustering
Filtering best features.
SIFT matching
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
First approaches:
1.
2.
3.
4.
5.
SURF matching.
Template database SURF matching.
Clustering
Filtering best features.
SIFT matching
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
1. SURF matching:
• Speeded up robust features.
• Simplification of SIFT.
• Fast.
• “Combination of novel detection,
description, and matching steps.” [ETH
Swiss Federal Institute of Technlogy]
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
1. SURF matching
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
1. SURF matching
Result with a rigid feature:
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
1. SURF matching
Result (no more fancy glasses):
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
First approaches:
1.
2.
3.
4.
5.
SURF matching.
Template database SURF matching.
Clustering
Filtering best features.
SIFT matching.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
2. Template database SURF matching.
Test image
Classifier
(NN)
Output image
with matchings
FACE/NON FACE
templates
(SURF
descriptors)
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
2. Template database SURF matching
+
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
2. Template database SURF matching
Result A
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
2. Template database SURF matching
Result B
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
First approaches:
1.
2.
3.
4.
5.
SURF matching.
Template database SURF matching.
Clustering
Filtering best features.
SIFT matching.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
3. Clustering:
Test image
Classifier
(NN)
FACE/NON FACE
Clusters of SURF
descriptors
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
Output image
with matchings
Clustering!
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
3. Clustering:
Ideal
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
Reality
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
First approaches:
1.
2.
3.
4.
5.
SURF matching.
Template database SURF matching.
Clustering
Filtering best features.
SIFT matching.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
4. Filtering best features:
• Find the extreme ratios.
• Two separate databases.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
4. Filtering best features:
Blue BAD, Orange GOOD
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
SURF FAILS
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
SURF FAILS
But...WHY?
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
SURF FAILS
[REF] [A Comparison of SIFT, PCA-SIFT and SURF - Luo Juan,
Oubong Gwun- International Journal of Image Processing2010]
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
First approaches:
1.
2.
3.
4.
5.
SURF matching.
Template database SURF matching.
Clustering
Filtering best features.
SIFT matching.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
5. SIFT Matching:
Test image
Classifier
(NN)
FACE/NON FACE
Clusters of SIFT
descriptors
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
Output image
with matchings
SIFT!
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
5. SIFT Matching:
• Bringing in SIFT.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Exploration into local invariant features
5. SIFT Matching:
• We should do anything else...
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Contents
• Introduction
• Exploration into local invariant features
• The final method
• Results
• Conclusions
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
The final method
1. Skin detection:
• Help SIFT as much as possible.
• Make the ROI of skin color regions.
• Methodology:
• Extract skin segments: find the mean
and covariance between CrCb.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
The final method
1. Skin detection:
• Methodology:
• Obtain the Gaussian probabilities of
skin color pixels for the input image.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
The final method
2. Skin detection:
• Methodology:
• Apply an adaptive threshold.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
The final method
2. Skin detection:
• Methodology:
• Avoid region without holes and apply
morphological operation over the
binary image.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
The final method
3. Face detection:
• Filtering best features.
• Finding matches between each skin
region and each template.
• D.Lowe
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
The final method
3. Face detection:
• Check if each skin region is a face.
• Threshold = Face matches / No face
matches
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Contents
• Introduction
• Exploration into local invariant features
• The final method
• Results
• Conclusions
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Results
ans =
TestImage_num: 1
Regioin_num: 1
Keypoint_num: 1135
TotalFaceMatch_num: 99
TotalNoFaceMatch_num: 14
TemplateMatchNum: [2x37 double]
ans =
TestImage_num: 1
Regioin_num: 2
Keypoint_num: 74
TotalFaceMatch_num: 3
TotalNoFaceMatch_num: 2
TemplateMatchNum: [2x37 double]
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Results
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Results
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Results
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Contents
• Introduction
• Exploration into local invariant features
• The final method
• Results
• Conclusions
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Conclusions
• We have implement a face-detection method
based on feature descriptors with a lot of help.
• SIFT over SURF
• QUALITY OF THE FEATURES over QUANTITY
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez
Conclusions
•How can be improved?
• Use shape characteristics.
• A larger template database and a more
robust classifier.
• Use morphology advantages.
• Apply other methods instead of feature
description (i.e: HAAR).
• Use previous frames information.
Master Erasmus Mundus of Science in Vision and Robotics (VIBOT)
- X.Huang, M.Ison, D.Martínez