Nov 3rd 2013 – IROS 2013 – Cloud Robotics Workshop
Life-long Learning Perceptionusing Cloud Database Technology
Tim Niemueller, Stefan Schiffer, and Gerhard LakemeyerKnowledge-based Systems Group, RWTH Aachen University
Safoura Rezapour-LakaniIntelligent and Interactive Systems, University of Innsbruck
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Motivation
Identify specific objects of interest in sensor range.
Describe objects by attributes instead of classesLearn objects over timeIntegrate new perception approaches over timeAccommodate various sensor modalities
Distributed Robot Perception DatabaseExtend perception system and object database over time
and share it among robots.
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 1 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Motivation
Identify specific objects of interest in sensor range.
Describe objects by attributes instead of classesLearn objects over timeIntegrate new perception approaches over timeAccommodate various sensor modalities
Distributed Robot Perception DatabaseExtend perception system and object database over time
and share it among robots.
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 1 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Perception Classifier Cascades
Find a red apple.
Candidates
apple color:redYes Yes
NoMatches?
NoResult
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 2 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Architecture
Classifiers
Base
Attribute
Meta
Data
Objects
Descriptors
Attributes
Queries
. . .
SURF Color · · · Haar VFH Shape
pepper color:red color:yellow apple . . .
Query: {color : red, apple}
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 3 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Perception Database Requirements
Flexible Data Structuresvarying/evolving data structures
Data Managementunified storage architecturereplication, backup and restore
Flexible and Efficient Retrievalquery for specific datalow-overhead retrieval of diverse and large data
MongoDB, the document-oriented,schema-less database, is particularly
well-suited to fulfill these criteria.
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 4 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Perception Database Requirements
Flexible Data Structuresvarying/evolving data structures
Data Managementunified storage architecturereplication, backup and restore
Flexible and Efficient Retrievalquery for specific datalow-overhead retrieval of diverse and large data
MongoDB, the document-oriented,schema-less database, is particularly
well-suited to fulfill these criteria.
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 4 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
MongoDB and its Building Blocks
Document-orientedGrouped key-value pairs
Schema-lessNo declaration or enforcement ofparticular structure by DB
CollectionsSimilarly structured documentsIndexing reference frame
QueriesJavaScript based query languageSelect based on document fields
{ // attributes/classifiers// for specific object"_id" : ObjectId("50e..."),"data_id" : "apple_1_1_10_c""scene_id" : "apple_1_1_10""attributes" : {"apple" : true,"color" : "red"
}"classifiers" :["SIFT", "SURF", "Gabor","Haar", "Color", "VFH" ...]
}
{ // classifier info excerpt// for attribute doc"_id" : ObjectId("52..."),"data_id" : "apple_1_1_10_c""VFH" : {"model_file" :
"apple_1_1_10_c_vfh.txt""extract_time" : 20
},// [...]
}
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 5 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
MongoDB in our Perception System
DataSharing
Scenes Objects Raw Data
Classifiers
Attributes
Scene
Soda
Apple
Cup
Image, Point
Cloud, ...
SIFT, VFH, ...
attributes:apple, color:red
classifiers:sift, vfh, ...
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 6 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Query Example
Candidates
apple color:redYes Yes
NoMatches?
No Result
Q = {color : red, cup}
{attributes: {
"apple": true,"color": "red"
}}
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 7 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Query Example
Candidates
apple color:red
Yes Yes
NoMatches?
No Result
CDq = D.classifiers,
where q ∈ D.attributes
Cq =⋂
CDq 6=∅
CDq
docs = db.attributes.aggregate({$match: { "attributes.color": "red"}},{$project: {classifiers: 1}},{$group: {_id: "$classifiers"}})
}
set_intersect(docs, "classifiers");
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 7 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Query Example
Candidates
apple color:redYes Yes
NoMatches?
No Result
Oq = Cq(input)
O =⋂
q∈QOq
Apply classifiersFilter through cascade
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 7 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Why a Cloud Database?
Flexible Data StructuresEasily accommodate various data typesQuickly grow w/o tedious specification
Query CapabilitiesFormulate queries deep into data structuresMajor benefit over traditional file system storage
Distributed DataReplicate for off-line trainingQuick boot-strapping for new robots/methodsSharding for multi-host robots
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 8 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Full Retraining
0
50
100
150
200
250
300
350
400
450
500
0 500 1000 1500 2000
Tim
e [m
s]
# of iterations
DB write
DB read
SIFT
SURF
Gabor
Shape
Color
VFH
Cylinder
Sphere
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 9 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Full Retraining
0
50
100
150
200
250
300
350
400
450
500
0 500 1000 1500 2000
Tim
e [m
s]
# of iterations
DB write
DB read
SIFT
SURF
Gabor
Shape
Color
VFH
Cylinder
Sphere
0
50
100
150
200
250
300
350
400
450
500
0 500 1000 1500 2000
Tim
e [m
s]
# of iterations
DB write
DB read
SIFT
SURF
Gabor
Shape
Color
VFH
Cylinder
Sphere
Gabor training magnified 4xTime [ms]
50
DB Read
DB Write
Training
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 9 / 10
Life-long Learning Perception using Cloud Database TechnologyMotivation Perception Database Evaluation Conclusion
Conclusion and Questions
Cloud database for perception which is extensiblein terms of known objects and perception methods.
Attribute-based perceptionIncreasing number of objectsEvolve perception methods
Flexible storage w/ MongoDBCapable query featuresShare data among robots
www.fawkesrobotics.org
Niemueller, Schiffer, Lakemeyer, Lakani Nov 3rd 2013 @ IROS 2013 10 / 10