Articulated Hand Posture Recognition System using IDSC Features - - PowerPoint PPT Presentation

articulated hand posture recognition system using idsc
SMART_READER_LITE
LIVE PREVIEW

Articulated Hand Posture Recognition System using IDSC Features - - PowerPoint PPT Presentation

Articulated Hand Posture Recognition System using IDSC Features Sourav Khandelwal CS676: Computer Vision and Image Processing Indian Institute of Technology, Kanpur April 12, 2011 Sourav Khandelwal (CS676: Computer Vision and Image


slide-1
SLIDE 1

Articulated Hand Posture Recognition System using IDSC Features

Sourav Khandelwal

CS676: Computer Vision and Image Processing Indian Institute of Technology, Kanpur

April 12, 2011

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 1 / 16

slide-2
SLIDE 2

Outline

1

Problem Statement

2

Inner Distance Shape Context(IDSC)

3

Methodology

4

IDSC Descriptors

5

Experiments

6

Questions

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 2 / 16

slide-3
SLIDE 3

Problem Statement

Problem Statement

To capture and recognize various articulated hand postures using inner-distance shape context(IDSC) descriptor. Hand Postures

Figure: Various hand postures: Hand states(open/closed), Sign Languages , Grasp Patterns

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 3 / 16

slide-4
SLIDE 4

Problem Statement

Two problems

(a) Two State Classification(open/closed) (b) Sign Language Classification

Challenging Problem Complexity of Hand Articulations: 27 DOFs. The occlutions of fingers.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 4 / 16

slide-5
SLIDE 5

Problem Statement

Two problems

(c) Two State Classification(open/closed) (d) Sign Language Classification

Challenging Problem Complexity of Hand Articulations: 27 DOFs. The occlutions of fingers.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 4 / 16

slide-6
SLIDE 6

Inner Distance Shape Context(IDSC)

Shape Context

The shape context is a descriptor used to measure similarity and point correspondences between shapes. It describes each point along the objects’ contour wrt to other points in the contour. The descriptor of a point pi is a coarse histogram hi of the relative coordinates of remaining n − 1 points. hi(k) = #{q = pi : (q − pi) ∈ bin(k)} The computation of histogram is based on both relative distance and angle. The Shape Context descriptor is a robust, compact and highly discriminative description of objects as it captures the distribution of each point relative to all

  • ther points.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 5 / 16

slide-7
SLIDE 7

Inner Distance Shape Context(IDSC)

Inner Distance Shape Context

The inner distance is proposed by Haibin Ling. Proved to be very useful for articulated objects. Normal Shape context uses L2 distance measure. Whereas, the IDSC descriptor uses the inner distance calculated as the length of the shortest path within the shape boundary. Invariant to shape articulations.

Figure: inner distance using points on shape boundary

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 6 / 16

slide-8
SLIDE 8

Inner Distance Shape Context(IDSC)

Inner Distance Shape Context

The inner distance is proposed by Haibin Ling. Proved to be very useful for articulated objects. Normal Shape context uses L2 distance measure. Whereas, the IDSC descriptor uses the inner distance calculated as the length of the shortest path within the shape boundary. Invariant to shape articulations.

Figure: inner distance using points on shape boundary

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 6 / 16

slide-9
SLIDE 9

Methodology

Methodology

The palm regions are extracted from hand images using thresholding and mean smoothing. The longest contour of the image is extracted from the image boundary. 200 points are taken uniformly from the contour and represented as a graph. IDSC descriptor is calculated using these 200 points. This returns a histogram description of each point along the objects’ contour to describe other points in the contour wrt to distance and angle.. These signatures are used as feature representation for input in SVM to address the classification problem.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 7 / 16

slide-10
SLIDE 10

Methodology

Computation of IDSC Descriptor

Inner-distance computation The idea is to find out the length of the shortest path between two points on the shape. A graph is built using sampled points.For each pair of sample points p1 and p2 , if the line segment connecting p1 and p2 falls entirely within the object, an edge between p1 and p2 is added to the graph with its weight equal to the Euclidean distance ||p1 − p2||. All pair shortest distance matrix is computed using Bellman-Ford’s all pair shortest path algorithm. Relative distance between points is described by inner distance. Inner-angle computation Relative orientation between points p and q is calculated as the angle between contour tangent at p and the direction of path (p,q). This is called inner-angle. The inner angle is calculated along calculation of shortest distance between all points.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 8 / 16

slide-11
SLIDE 11

Methodology

Computation of IDSC Descriptor

Inner-distance computation The idea is to find out the length of the shortest path between two points on the shape. A graph is built using sampled points.For each pair of sample points p1 and p2 , if the line segment connecting p1 and p2 falls entirely within the object, an edge between p1 and p2 is added to the graph with its weight equal to the Euclidean distance ||p1 − p2||. All pair shortest distance matrix is computed using Bellman-Ford’s all pair shortest path algorithm. Relative distance between points is described by inner distance. Inner-angle computation Relative orientation between points p and q is calculated as the angle between contour tangent at p and the direction of path (p,q). This is called inner-angle. The inner angle is calculated along calculation of shortest distance between all points.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 8 / 16

slide-12
SLIDE 12

Methodology

Computation of IDSC Descriptor

Inner-distance computation The idea is to find out the length of the shortest path between two points on the shape. A graph is built using sampled points.For each pair of sample points p1 and p2 , if the line segment connecting p1 and p2 falls entirely within the object, an edge between p1 and p2 is added to the graph with its weight equal to the Euclidean distance ||p1 − p2||. All pair shortest distance matrix is computed using Bellman-Ford’s all pair shortest path algorithm. Relative distance between points is described by inner distance. Inner-angle computation Relative orientation between points p and q is calculated as the angle between contour tangent at p and the direction of path (p,q). This is called inner-angle. The inner angle is calculated along calculation of shortest distance between all points.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 8 / 16

slide-13
SLIDE 13

IDSC Descriptors

IDSC

IDSC Computation The histogram is computed based on both inner-distance and inner-angle for each point on the contour. This descriptor is used as a feature representation for a particular hand posture image.

Figure: inner distance using points on shape boundary

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 9 / 16

slide-14
SLIDE 14

IDSC Descriptors

IDSC

IDSC Computation The histogram is computed based on both inner-distance and inner-angle for each point on the contour. This descriptor is used as a feature representation for a particular hand posture image.

Figure: inner distance using points on shape boundary

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 9 / 16

slide-15
SLIDE 15

IDSC Descriptors

IDSC

IDSC Computation The histogram is computed based on both inner-distance and inner-angle for each point on the contour. This descriptor is used as a feature representation for a particular hand posture image.

Figure: inner distance using points on shape boundary

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 9 / 16

slide-16
SLIDE 16

IDSC Descriptors

Visualizing IDSC

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 10 / 16

slide-17
SLIDE 17

Experiments

Experiments

Hand State Recognition The SVM is trained using the IDSC descriptions of open/closed hand contour images, 100 examples per state. The hand postures contained very high in−plane rotations(upto +/ − 180) and substantial out−of−plane rotations(upto +/ − 45). The training set is built such that it has images from all the rotations. The algorithm is tested on 100 images each of closed and open states. Results 64% recognition with normal SC descriptors. 74% recognition with IDSC descriptors. These results were improved after including a pre−processing step.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 11 / 16

slide-18
SLIDE 18

Experiments

Experiments

Hand State Recognition The SVM is trained using the IDSC descriptions of open/closed hand contour images, 100 examples per state. The hand postures contained very high in−plane rotations(upto +/ − 180) and substantial out−of−plane rotations(upto +/ − 45). The training set is built such that it has images from all the rotations. The algorithm is tested on 100 images each of closed and open states. Results 64% recognition with normal SC descriptors. 74% recognition with IDSC descriptors. These results were improved after including a pre−processing step.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 11 / 16

slide-19
SLIDE 19

Experiments

Experiments

Hand State Recognition The SVM is trained using the IDSC descriptions of open/closed hand contour images, 100 examples per state. The hand postures contained very high in−plane rotations(upto +/ − 180) and substantial out−of−plane rotations(upto +/ − 45). The training set is built such that it has images from all the rotations. The algorithm is tested on 100 images each of closed and open states. Results 64% recognition with normal SC descriptors. 74% recognition with IDSC descriptors. These results were improved after including a pre−processing step.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 11 / 16

slide-20
SLIDE 20

Experiments

Pre−Processing

The training and test images are grouped into 10 different bins uniformly based on their primary orientation direction. The orientation direction is estimated by computing the scatter direction of the images through principal component analysis (PCA). The SVM is trained with the hand state instances of the corresponding orientation. To test, the primary direction of the test image is computed and is projected on the appropriate SVM for classification. The results improved upto 81% recognition.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 12 / 16

slide-21
SLIDE 21

Experiments

Sign Language Recognition

Not yet done. Planning to work on 5 different sign language and use leave-one-out cross-validation technique to classify.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 13 / 16

slide-22
SLIDE 22

Experiments

Datasets

Hand Image Dataset from Image and Video computing group, Boston University. The link to the database is here. Own database of hand posture images of 2−3 different persons for testing purposes.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 14 / 16

slide-23
SLIDE 23

Experiments

References

Gopalan, Raghuraman and Dariush, Behzad, Toward a vision based hand gesture interface for robotic grasping, IEEE/RSJ international conference on Intelligent robots and systems, pages 1452–1459, 2009.

  • H. Ling and D.W. Jacobs. Shape Classification Using the Inner-Distance. IEEE
  • Trans. Pattern Analysis and Machine Intelligence (PAMI), pages 286-299, 2007.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 15 / 16

slide-24
SLIDE 24

Experiments

References

Gopalan, Raghuraman and Dariush, Behzad, Toward a vision based hand gesture interface for robotic grasping, IEEE/RSJ international conference on Intelligent robots and systems, pages 1452–1459, 2009.

  • H. Ling and D.W. Jacobs. Shape Classification Using the Inner-Distance. IEEE
  • Trans. Pattern Analysis and Machine Intelligence (PAMI), pages 286-299, 2007.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 15 / 16

slide-25
SLIDE 25

Experiments

References

Gopalan, Raghuraman and Dariush, Behzad, Toward a vision based hand gesture interface for robotic grasping, IEEE/RSJ international conference on Intelligent robots and systems, pages 1452–1459, 2009.

  • H. Ling and D.W. Jacobs. Shape Classification Using the Inner-Distance. IEEE
  • Trans. Pattern Analysis and Machine Intelligence (PAMI), pages 286-299, 2007.

Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 15 / 16

slide-26
SLIDE 26

Questions Sourav Khandelwal (CS676: Computer Vision and Image ProcessingIndian Institute of Technology, Kanpur) Hand Posture Recognition April 12, 2011 16 / 16