Using Depth and Appearance Features for Informed Robot ... · Using Depth and Appearance Features for Informed Robot Grasping of Highly Wrinkled Clothes Arnau Ramisa XXXIV Jornadas

Using Depth and Appearance Features for

Informed Robot Grasping of Highly Wrinkled

Clothes

Arnau Ramisa

XXXIV Jornadas de Automática

September, 2013

Introduction

Introduction

Grasping:

Detection of 3D objects models

Introduction

Grasping:


Rigid objects: pose (6 DoF)

Introduction

Grasping:


Rigid objects: pose (6 DoF)Textile objects: pose+deformation

Introduction

Grasping:



Current approaches use multiple re-grasps

Image credit: J. Maitin-Shepard et al. "Cloth grasp point detection based on multiple-view geometric cues with

application to robotic towel folding", ICRA10

Introduction

Grasping:



Current approaches use multiple re-grasps

Our contribution: computer vision basedapproach to shorten manipulation timeusing depth and appearance information

Proposed Method - Appearance and depth features

Appearance: Scale Invariant Feature

Transform (SIFT)

Depth: Geodesic Depth Histogram

(GDH). Its original names is Geodesic

Intensity Histogram (GIH)1, but we

apply it to depth data

1: Ling and Jacobs, Deformation Invariant Image Matching, ICCV05

Proposed Method - Bag of visual and depth words


1 Local feature extraction(SIFT and GDH)



2 Quantize the features intovisual words




3 For each bounding box:

























4 Accumulate visual wordsin a histogram






5 Classify the histogramwith logistic regression




















6 Responses from all windows arecombined in a probability map



7 Maxima of this probability mapare good collar hypotheses




8 Refinement: χ2 RBF SVM inarea surrounding initialhypotheses









9 Grasp point selected at centerof detection bounding box





9 Grasp point selected at centerof detection bounding box

10 In real experiments: Goodgrasping point is found using agraspability measure

Evaluation - Polo dataset

Dataset of polos and other clothes

Manually annotated ground truthbounding boxes

Acquired with a Kinect camera

Divided in three subsets

Blue polo: 70% train 30% testOthers: Polos and shirts otherthan the blue poloMixed: Polos mixed with pants,t-shirts, etc.

Evaluation criterion: Collarcorrectly detected if final grasppoint lies inside the ground truthbounding box






















Evaluation - True positive ratio per image

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

blue polo

tpr@1 tpr@2 tpr@3 tpr@4 tpr@5

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7others

SIFT K64SIFT K128

GDH K16GDH K32

GDH K16SIFT K64

GDH K16SIFT K128

GDH K32SIFT K64

GDH K32SIFT K128

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45mixed

Evaluation - Blue polo subset

117 images. 30% test/70% train

0.0 0.2 0.4 0.6 0.8 1.0

Recall

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Pre

cis

ion

SIFT k64 (ap=0.496)

SIFT k128 (ap=0.572)

GDH k16 (ap=0.149)

GDH k32 (ap=0.187)

GDH k16 SIFT k64 (ap=0.453)



GDH k32 SIFT k128 (ap=0.463)

Evaluation - Others subset

48 images. Other polos and shirts

0.0 0.2 0.4 0.6 0.8 1.0

Recall

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Pre

cis

ion

SIFT k64 (ap=0.262)

SIFT k128 (ap=0.331)

GDH k16 (ap=0.136)

GDH k32 (ap=0.149)


GDH k16 SIFT k128 (ap=0.316)



Evaluation - Mixed subset

29 images (48 collars annotated). Polos mixed withother clothes

0.0 0.2 0.4 0.6 0.8 1.0

Recall

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Pre

cis

ion

SIFT k64 (ap=0.254)

SIFT k128 (ap=0.249)

GDH k16 (ap=0.063)

GDH k32 (ap=0.077)


GDH k16 SIFT k128 (ap=0.166)


GDH k32 SIFT k128 (ap=0.165)

Recap

Pros:

Exploiting computer vision to better model textile objects∼ 70% correct detections with the top detection in each imageDepth information helps the method generalize

Cons:

Geodesic Depth Histogram descriptors alone perform poorlyAND they are very expensive to compute (+1 hour/image)

Work published in: Ramisa et al. “Using depth and appearance features for

informed robot grasping of highly wrinkled clothes”, ICRA, 2012.

The FINDDD descriptor

Most current 3D descriptors are expensive to compute



Searching neighbors in an unstructured point cloud



Searching neighbors in an unstructured point cloudComputing an invariant reference frame




In a garment manipulation scenario things can be simplified





Actions are intrinsically coupled to viewpoint





Actions are intrinsically coupled to viewpointUsing a range image (aka structured point cloud) allows usingIntegral Images





Actions are intrinsically coupled to viewpointUsing a range image (aka structured point cloud) allows usingIntegral Images

We propose new depth descriptor that takes advantage of thesesimplifications to be two orders of magnitude faster





Best representation of the normals to define bin boundaries?



Spherical coordinates: bins defined as intervals in φ and θ

(φ, θ) =(

arccos(z

r

)

, arctan(y

x

))


Best representation of the normals to define bin boundaries?Spherical coordinates: bins defined as intervals in φ and θ

(φ, θ) =(

arccos(z

r

)

, arctan(y

x

))


Best representation of the normals to define bin boundaries?Spherical coordinates: bins defined as intervals in φ and θ

Cartesian coordinates: quasi-equidistant bins defined bytessellating a half-sphere





Soft voting used to gain robustness to aliasing





Soft voting used to gain robustness to aliasing

Integral Images make cost independent of local region size

20 40 60 80 100 120patch side size

0

10

20

30

40

50

60

70

80

compu

tatio

n tim

e (sec

)

Integral ImagesNo Integral Images

Σ = A + B − C − D

The FINDDD descriptor: Results

Computational time

Descriptor Time (s) Time per desc. (ms)

FINDDD (B=13) 4.0 0.015FINDDD (B=41) 10.5 0.040SHOT 482.5 1.98FPFH 313.5 1.28

Evaluated on a 3Ghz Linux machine. Result obtained extracting a descriptor for every

point in a 640 × 480 structured point cloud (average of 10 runs).


Wrinkle Retrieval

B S SV mAPFINDDD 13 21 T 59.8

13 21 F 52.213 43 T 80.513 43 F 80.613 65 T 85.913 65 F 86.1

41 21 T 61.241 21 F 57.141 43 T 82.241 43 F 77.641 65 T 86.041 65 F 81.8

FPFH 62.0SHOT 50.6

B: Number of binsS: Local area sizeSV: Soft voting


Garment Recognition

RBF-χ2 SVM — mAPGarment FINDDD SHOT FPFHDress 66.8 61.9 67.6

Shirt 54.5 72.9 79.7

T-Shirt 84.7 70.1 76.5Jeans 72.9 65.1 77.9

Polo 96.0 83.7 77.6Sweater 84.6 92.1 93.7

Average 76.6 74.3 78.8


Informed Grasping Point Selection (real experiment)Correct grasping point (polo lapel) accuracy: 82.6% (23 rep)

Recap

Pros:Novel shape descriptor for textile manipulationCompetitive performance with competing approachesFast computation time thanks to simplifications

Cons:Reduced invariance to viewpointNot robust to strong depth discontinuities

Work published in: Ramisa et al. “FINDDD: A Fast 3D Descriptor to

Characterize Textiles for Robot Manipulation”, IROS, 2013.

ROBOCUP13

Proposed method used in Robocup@Home’13 by Reem@IRI team

ROBOCUP13

Robocup@Home is a competition for domestic robots(Reem@IRI finished 7 out of 21 teams)

Tasks include person following and interaction, objectrecognition and grasping or emergency situation detection(fire and smoke)

Also an Open Challenge to demonstrate unique abilities of therobot

Reem grasped and hung a polo in a hanger (worth ∼ 42% ofits total score)

ROBOCUP13

Conclusions

Thank You!

Questions?

Using Depth and Appearance Features for Informed Robot ... · Using Depth and Appearance Features for Informed Robot Grasping of Highly Wrinkled Clothes Arnau Ramisa XXXIV Jornadas

Documents