26
Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University of Massachusetts Boston

Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Embed Size (px)

Citation preview

Page 1: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Cascaded Classifier for Automatic Crater Detection

Henry Z. Lo

Advisor: Wei DingDomain Scientist: Tomasz Stepinski

Knowledge Discovery LabUniversity of Massachusetts Boston

Page 2: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Overview

• Introduction:o Cascading classifier.o Experimental road map.

• Experiments:o Tests on feature sets.o Tests on positive example training set content.o Tests on negative example training set size.o Tests on negative example training set content. 

• Discussion:o Implications of results.o Unresolved issues.o Future directions.

Page 3: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Cascading Classifier

• Architecture:o Layers of Adaboost classifiers.o Each layer trained on the FP of previous layer. o Input must be accepted by all, sequentially, to be

considered a crater.o Rejection can happen at any stage.

 

Page 4: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Cascading Classifier

• Features:o Exclusively uses Haar-like features.o Can be calculated in constant time.o Contrast based.o Scanned over entire subwindow.

 

Page 5: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Cascading Classifier

• Implementation:o Used OpenCV implementation.o Free and open source. o Many variables:

Number of layers. "Minimum hit rate" - false positive rate. "Max false alarm" - false negative rate. 3 feature sets.

 

Page 6: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Experimental Road Map

• Tweak for performance:o OpenCV parameters.o Features.o Training set.

• The following OpenCV parameters improve performance:o Minimum hit rate.o Max false alarm.o Number of layers.

• Still need to tweak features and training sets for:o Training time.o Generalizability.

• L 

Page 7: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Experimental Road Map

Page 8: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

• Each of these factors will be tested individually for effect on precision, recall, and F1.

  • We avoid studying interaction effects for simplicity.

 • In the future, we will investigate how to combine different

features and test sets for optimal result.

Experimental Road Map

Page 9: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

• We use tile 3-24 for both training and testing.• This tile was chosen for its relatively smooth surface.• Future studies will test on other tiles as well.

 

Experimental Road Map

Page 10: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Feature Set Variation

Page 11: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Feature Set Variation

• OpenCV offers 3 different feature sets: o CORE:      1a, 1b, 2a, 2c.o BASIC:      CORE + 2b, 2d, 3ao ALL:          all features

 • Since ALL is a superset of CORE and BASIC, it should

perform best.

Page 12: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Feature Set Variation

• In recall, CORE and BASIC outperformed ALL.

• In precision and F1, the exact opposite was true.

Page 13: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Haar Features

Page 14: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Haar Features

• Inclusion of tilted features beneficial to performance. • More features than those given may provide further benefit.

 • It is not obvious how to create Haar features in OpenCV.

 • Postponing creation of specialized Haar features.

Page 15: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Ground Truth Windows

Page 16: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Ground Truth Windows

• Positive examples contained tightly cropped craters. • No crater rims or surrounding area.

 • Experimented with including area around craters. •  • Range: 1x crater radius - 2x crater radius, in steps of .1.

              1.0    1.2     1.4       1.6       1.8         2.0

Page 17: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Ground Truth Windows

• As the subwindow increased, precision and F1 increased.

 • However, recall suffered.

Page 18: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Negative Example Set Size

Page 19: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Negative Example Set Size

• All classifiers tested were trained on 300 negative examples.  • By providing the classifier with more negative examples, we

give it more information. • Performance should increase with more negative examples.

 • Tested classifiers trained on 300, 400, 500, 600, and 700

negative examples.

Page 20: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Negative Example Set Size

• F1 and precision increase with more negative examples.

• Recall decreases.

Page 21: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Negative Example Manipulation

Page 22: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Negative Example Manipulation

• The idea is to put some false positives back into the training set.

  • This will teach the classifier using its own mistakes.

 • However, selecting the false positives is rather difficult, as

we will see later.

Page 23: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Result Implications

• Window scaling has the most noticeable effect on F1, recall, and precision.

 • Next most important is the feature set used.

 • The number of negative training examples is the least

important; however, this may be due to the small range of values being tested.

Page 24: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Future Directions

 • Once optimal features and training sets are found, we can

manipulate OpenCV variables.  • Recall that the classifier may be improved by the following:

o More layers in the classifier. 

o Setting the minimum hit rate (recall).•  

o Setting the max false alarm rate (precision).•  • Time complexity of classifier training requires further study.

   

Page 25: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Future Directions

• Further exploration of cascaded classification algorithm: 

o Testing classifier on other tiles. • Exploration of other object detection algorithms.

 o Neural networks.

Page 26: Cascaded Classifier for Automatic Crater Detection Henry Z. Lo Advisor: Wei Ding Domain Scientist: Tomasz Stepinski Knowledge Discovery Lab University

Questions?