Shape Matching And Object Recognition

Seminar On CSE-4102

SHAPE MATCHING AND OBJECT RECOGNITION USING

SHAPE CONTEXTS

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

1

ØIt is easy for human to make difference between two similar object. ØIt is difficult for machine to make difference between two similar object.

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

2

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

3

Shape Context : ØIt is Shape descriptor of shape matching.

Sample(a)

Sample(b)

that play the role

Log polar histogram

Correspond found using bipartite matching

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

4

Bipartite graph matching : ØIf cij denotes the cost between two point the cost is determined by: Where,

pi is a point on the first shape. ( shape ( a ) ) . pj is a point on the second shape. ( shape ( b ) ) . Ø

ØT h e c o n c e p t o f u s i n g d u m m y n o d e . To minimize Total cost. ØTotal cost Qudrat-E-Alahy of matching : Ratul, KUET, Khulna, Bangladesh 5

Idle state : ØWe use affine model to choose a suitable family of transformation. ØA standard choice of affine model: T(x)=Ax+o ØWe use TPS(Thin Plate Spline) model transformation. Ø

Regularization :

ØIf there is noise in specified values then the interpolation is relaxed by regularization. ØRegularization parameter determine the amount of smoothing . Ratul, KUET, Khulna, Bangladesh Qudrat-E-Alahy 6

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

7

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

8

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

9

Digit recognation : Error is only 63 %

using 20,000 training example.

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

10

3 - D object detection : Using 72 view per object.

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

11

proach is estimation of shape similarities and corresponden

picture is used.

hile experimenting.

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

12

Thank you

Qudrat-E-Alahy Ratul, KUET, Khulna, Bangladesh

13