Unformatted text preview:

Class projects EE148 Joseph Gonzales Sue Ann Hong Theodore Yu Neil Halelamien Yang Cheng Fei Wang Chunhui Gu Ruxandra Paun Attractive Rating Attractive Rating Sharing Features YGP Dorko Schmid Viola and Jones Pictures with Words Viola and Jones Dorko Schmid Probabilistic individual object recognition P Moreels P Perona EE CNS 148 Spring 2005 California Institute of Technology The recognition continuum Individual objects BMW logo cars means of transportation variability Categories Lowe s recognition system Models Test image Lowe 99 04 The voting approach Hough transform x1 y1 s1 1 x2 y2 s2 2 Transform predicted by this match x x2 x1 y y2 y1 s s2 s1 2 1 Voting performed in the space of transform parameters s y x Lowe s recognition system Models Test image Lowe 99 04 Constellation model Burl 96 Weber 00 Fergus 03 Pros and Cons Lowe s recognition system Many parts redundancy Learn from 1 image Fast Manual tuning of parameters Rigid planar objects Sensitive to clutter Constellation model Principled detection recognition Learn parameters from data Model clutter occlusion distortions High number of parameters O n2 5 7 parts per model many training examples needed learning expensive How to adapt the constellation model to our needs Reducing degrees of freedom 1 Common reference frame Lowe 99 Huttenlocher 90 model m position of model m 2 Share parameters Schmid 97 3 Use prior information learned on foreground and background FeiFei 03 Parameters and priors Foreground Constellation model Clutter Gaussian Gaussian part Gaussian shape pdf Gaussian background appearance pdf relative scale pdf appearance pdf log scale Prob of detection 0 8 0 8 0 9 0 75 Foreground Sharing parameters Clutter Gaussian conditional Gaussian part shape pdf Gaussian Gaussian background appearance pdf relative scale pdf appearance pdf log scale Prob of detection Based on Fergus 03 Burl 98 0 2 0 8 0 2 0 2 Hypotheses features assignments New scene test image models from database Interpretation Hypotheses model position New scene test image Models from database 1 2 3 affine transformation Score of a hypothesis observed features geometry appearance Hypothesis model position assignments database of models Bayes rule Consistency constant Hypothesis probability Score of a hypothesis Consistency between observations and hypothesis foreground features geometry appearance null assignments geometry Probability of number of clutter detections Probability of detecting the indicated model features Prior on the pose of the given model appearance Efficient matching process Scheduling inspired from A scene features no assignment done empty hypothesis null assignment 1 assignment 2 assignments P P P P Score P P can be compared P P explore most promising branches first Pearl 84 Grimson 87 P P P P perfect completion admissible heuristic used as a guide for the search Increase computational efficiency at each node searches only a fixed number of sub branches forces termination Recognition the first match No clue regarding geometry first match based on appearance st be re featu tch a m Initialization of hypotheses queue s se PPPPP match t s e b cond PPPPP New scene models from database PPPPP Scheduling promising branches first re featu Updated hypotheses queue s PPPP models from database New scene PPP PPP Experiments Toys database models 153 model images Toys database test images scenes 90 test images multiple objects or different view of model Kitchen database models 100 model images Kitchen database test images 80 test images 0 9 models test image Lowe s method Our system Examples Test image Test image Identified model Lowe s model implemented using Lowe 97 99 01 03 Identified model Performance evaluation a Object found correct pose Detection Test image hand labeled before the experiments b Object found incorrect pose False alarm c Wrong object found False alarm d Object not found Non detection Models database Scenes test images Results Toys images 153 model images 90 test images 0 5 models test image 80 recognition with false alarms test set 0 2 Lower false alarm rate than Lowe s system Results Kitchen images 100 training images 80 test images 0 9 models test image 254 objects to be detected Achieves 77 recognition rate with 0 false alarms Conclusions Unified treatment Best of both worlds Probabilistic interpretation of Lowe 99 04 Extension of Burl Weber Fergus 96 03 to many features many models one shot learning Higher performance Comparison with Lowe 99 04 Future work categories Back to the front end the features or interest points Moving the viewpoint Features stability Features stability is not perfect 240 keypoints extracted 232 keypoints extracted We want stable features video First stage feature detector difference of gaussians Crowley 84 Kadir Brady Kadir 02 Harris Harris 88 Affine invariant Harris Mikolajczyk 02 Traditional features detectors Distinctive points high X and Y gradients use of the square gradient matrix X I X I Y I X I 2 Y I X I Y I X I Y I 2 Y I X I If enough variability average of will have rank 2 Lucas Kanade maxima of smallest eigenvalue 2 Harris maxima of det 0 04 tr Foerstner maxima of det tr I X features detectors second order 1 1 2 1 2 1 Laplacian Prewitt Difference of gaussians Hessian 0 1 0 1 4 1 0 1 0 1 2 1 2 4 2 1 2 1 2XX I det 2 XY I 2XY I 2 YY I Second stage feature descriptor SIFT Steerable filters Lowe 04 Differential invariants Schmid 97 Freeman 91 Shape context Belongie 02 Dataset 100 3D objects Viewpoints 45 apart Viewpoints 45 apart Viewpoints 45 apart Viewpoints 45 apart How to compute the ground truth flat surfaces it s easy Click on 4 points compute the homography Ground truth Epipolar constraints Testing setup Unrelated images used to load the database of features Distance ratio Correct matches are highly distinctive lower ratio Incorrect correspondences are random correspondences low distinctiveness and ratio close to 1 Lowe 04 Detectors descriptors tested Detectors Harris Hessian Harris affine Hessian affine Difference of gaussians MSER Kadir Brady Descriptors SIFT steerable filters differential invariants shape context PCA SIFT dista nce No back grou nd i mag es Mah alan obis Results viewpoint change Results lighting scale changes Change in light result averaged over 3 lighting conditions Change in scale 7 0mm to 14 6mm 2D vs 3D Ranking of detectors descriptors combinations are modified when switching from 2D to 3D objects Conclusions of the comparisons Automated ground truth for 3D


View Full Document

CALTECH EE 148A - Lecture Notes

Download Lecture Notes
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Lecture Notes and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Lecture Notes 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?