autonomous learning of robust visual object detection & identification on a humanoid...
DESCRIPTION
n this work we introduce a technique for a hu- manoid robot to autonomously learn the representations of objects in its visual environment. Our approach involves feature- based segmentation of the images followed by learning to identify the object using Cartesian Genetic Programming. The learned identification is able to provide robust and fast segmentation of the objects, without using features. To allow for autonomous learning an attention mechanism is coupled with the training process. We showcase our system on a humanoid robot.TRANSCRIPT
#icdl/epirob 2012
autonomous learning of robust visual object
detection & identificationon a humanoid
S. Harding, P. ChandrashekhariahM. Frank, G. Spina,
A. Förster, J. Triesch, J. Schmidhuber
idsia / usi / supsi, machine intelligence, fias
Jürgen ’Juxi’ Leitner
manipulation
our iCubsetup is for
perceptionvisual
thanks to G. Metta and IIT for this picture
challengethe
IDSIA’s three
Harding et al., GPTP 2012,Leitner et al., ICDL 2012 Leitner et al., BICA 2012Leitner et al., IROS 2012
parts
cv approachescurrent
objectsdetecting
Harding et al., GPTP 2012
approachlearningour
cartesian genetic
programming
+ min dilate avg INP INP INP
+ min dilate avg INP INP INP
3""#2""#1"4.3"""""
Func,on"Connec,on"1"Connec,on"2"A"real"number" cartesian
genetic programming
detection
icImage GreenTeaBoxDetector::runFilter() { ! icImage node0 = InputImages[6];! icImage node1 = InputImages[1];! icImage node2 = node0.absdiff(node1);! icImage node5 = node2.SmoothBilateral(11);! icImage node12 = InputImages[0];! icImage node16 = node12.Sqrt();! icImage node33 = node16.erode(6);! icImage node34 = node33.log();! icImage node36 = node34.min(node5);! icImage node49 = node36.Normalize();
//cleanup ... icImage out = node49.threshold(230.7218f);! return out; }
detect
detect
approachsupervised learning
BUT:
segmentationfeature
saliencymap
collaborationFIAS
segmentationpre
approachcombined
MoBeEframework Frank et al., ICINCO, 2012.
salient object detection
conclusion
rough feature-based segmentationautomatic training set generation
+
=
cgp-based, robust filter-learningfor
for listeningthanks
[email protected] http://Juxi.net/projects http://robotics.idsia.ch
video at http://robotics.idsia.ch/im-clever/