- Home
- Documents
*Shape Context and Chamfer Matching in Cluttered Scenes Arasanathan Thayananthan Bj¶rn Stenger...*

prev

next

of 44

View

214Download

1

Tags:

Embed Size (px)

- Slide 1

Shape Context and Chamfer Matching in Cluttered Scenes Arasanathan Thayananthan Bjrn Stenger Dr. Phil Torr Prof. Roberto Cipolla Slide 2 What? Why? How? What: track articulated hand motion through video This work: tracker initialization Why: to drive 3D avatar (HCI) How: Using shape matching Two competing methods: chamfer matching shape context matching Slide 3 Goal: Hand Tracking [ICCV 03] Slide 4 How to detect a hand? Comparison of matching methods Shape context vs. Chamfer matching Enhancements for shape context Robustness to clutter Slide 5 Overview Shape Context matching in clutter Difficulties Proposed enhancements Comparison with original shape context Applications Hand detection EZ-Gimpy recognition Slide 6 Previous Work Shape Context [Belongie et al., 00] Invariance to translation and scale High performance in Digit recognition : MNIST dataset Silhouettes : MPEG-7 database Common household objects: COIL-20 database Chamfer Matching [Barrow et al., 77] efficient hierarchical matching [Borgefors, 88] pedestrian detection [Gavrila, 00] Slide 7 Shape Context: Histogram Shape context of a point: log-polar histogram of the relative positions of all other points Similar points on shapes have similar histograms Slide 8 Shape Context: Matching ij 2 Test Cost Function Bi-partite Graph Matching Optimal Correspondence Template Points Image Points Cost Matrix C Slide 9 Shape-Context: Matching Slide 10 Scale Invariance in Clutter ? Median of pairwise point distances is used as scale factor Clutter will affect this scale factor 50.541.6 50.5 121.9 Slide 11 Scale Invariant in Clutter ? Significant clutter Unreliable scale factor Incorrect correspondences Solution Calculate shape contexts at different scales and match at different scales Computationally expensive Slide 12 No Figural Continuity No continuity constraint Adjacent points in one shape are matched to distant points in the other Slide 13 Multiple Edge Orientations Edge pixels are divided into 8 groups based on orientation Shape contexts are calculated separately for each group Total matching score is obtained by adding individual 2 scores Slide 14 Single vs. Multiple Orientation Slide 15 Imposing Figural Continuity u i and u i-1 are neighboring points on the model shape u is the correspondence between two shape points Corresponding points v (i) and v (i-1) need to be neighboring points on target shape v u i-2 u i-1 uiui v (i-2) v (i) v (i-1) Slide 16 Imposing Figural Continuity u i-2 u i-1 uiui v (i-2) v (i) v (i-1) Slide 17 Imposing Figural Continuity Minimize the cost function for Ordering of the model shape is known Use Viterbi Algorithm Slide 18 With Figural Continuity Similar Shapes Slide 19 With Figural Continuity Different Scale Slide 20 With Figural Continuity Small Rotation Slide 21 With Figural Continuity Shape Variation Slide 22 With Figural Continuity Clutter Slide 23 Chamfer Matching Matching technique cost is integral along contour Distance transform of the Canny edge map Slide 24 Distance Transform Distance image gives the distance to the nearest edgel at every pixel in the image Calculated only once for each frame (x,y) d d Slide 25 Chamfer Matching Chamfer score is average nearest distance from template points to image points Nearest distances are readily obtained from the distance image Computationally inexpensive Slide 26 Chamfer Matching Distance image provides a smooth cost function Efficient searching techniques can be used to find correct template Slide 27 Chamfer Matching Slide 28 Slide 29 Slide 30 Slide 31 Slide 32 Multiple Edge Orientations Similar to Gavrila, edge pixels are divided into 8 groups based on orientation Distance transforms are calculated separately for each group Total matching score is obtained by adding individual chamfer scores Slide 33 Applications: Hand Detection Initializing a hand model for tracking Locate the hand in the image Adapt model parameters No skin color information used Hand is open and roughly fronto-parallel Slide 34 Results: Hand Detection Original Shape Context Shape Context with Continuity Constraint Chamfer Matching Slide 35 Results: Hand Detection Original Shape Context Shape Context with Continuity Constraint Chamfer Matching Slide 36 Applications: CAPTCHA Completely Automated Public Turing test to tell Computers and Humans Apart [Blum et al., 02] Used in e-mail sign up for Yahoo accounts Word recognition with shape variation and added noise Examples: Slide 37 EZ-Gimpy results 93.2% correct matches using 2 templates per letter Top 3 matches (dictionary 561 words) right 25.34 fight 27.88 night 28.42 Chamfer cost for each letter template Word matching cost: average chamfer cost + variance of distances Shape context 92.1% [Mori & Malik, 03] Slide 38 Discussion The original shape context matching Not invariant in clutter Iterative matching is used in the original shape context paper Correct point correspondence in the initial matching is quite small in substantial clutter Iterative matching will not improve the performance Slide 39 Discussion Shape Context with Continuity Constraint Includes contour continuity & curvature Robust to substantial amount of clutter Much better correspondences and model alignment just from initial matching No need for iteration More robust to small variations in scale, rotation and shape. Slide 40 Discussion Chamfer Matching Variant to scale and rotation More sensitive to small shape changes than shape context Need large number of template shapes But Robust to clutter Computationally cheap compared to shape context Slide 41 Conclusion Use shape context when There is not much clutter There are unknown shape variations from the templates (e.g. two different types of fish) Speed is not the priority Slide 42 Conclusion Chamfer matching is better when There is substantial clutter All expected shape variations are well- represented by the shape templates Robustness and speed are more important Slide 43 Forthcoming work [ICCV 2003] Slide 44 Webpage For more information on initialization and articulated hand tracking http://svr-www.eng.cam.ac.uk/~bdrs2/hand/hand.html