Download presentation
Presentation is loading. Please wait.
Published byConstance Quinn Modified over 8 years ago
1
Features-based Object Recognition P. Moreels, P. Perona California Institute of Technology
2
I : Individual object recognition Problem setup Coarse-to-fine hypothesis search. Probabilistic decomposition Experiments II : Searching for Grandma : Look at the eyes ! Problem setup Architecture of the system Demo Outline
3
The recognition continuum variability Individual objects Categories means of transportation BMW logo cars
4
Individual Object Recognition Autonomous navigation Security & Identification
5
… Problem setup New scene (test image) Models from database Find models and their pose (location, orientation…)
6
Models from database New scene (test image) Hypotheses – models & position 11 22 33 Θ = affine transformation variable H
7
Image characterization by features Features = privileged points of interest in an image Reduce the volume of information Interest points = high information content (e.g. high gradient) Hopefully provide invariance
8
Foreground vs. background features 1 2 3 4 5 6 7 8 Mutual euclidean distances in appearance space
9
Features assignments = models from database New scene (test image)... Interpretation... variable V
10
Scene interpretation scene model background
11
Coarse-to-fine example Fleuret & Geman (01) Face identification in complex scenes all scales, all tilts position in 8x8 cell all scales, right tilts position in 2x2 cell large scales, right tilts position in 2x2 cell
12
Coarse-to-fine strategy Search over possible locations is expensive Coarse-to-fine: progressive search refinement
13
Progressively narrow down focus on correct region of hypothesis space Use first information that is easy to obtain Simple building blocks organized in a cascade Probabilistic interpretation of each step Coarse-to-Fine detection
14
Prior knowledge Which objects are likely to be there, which pose are they likely to have ? unlikely situations
15
Coarse data: model voting Database Search tree model #10 model #12 model #25 model #12 Query features Candidate matches obtained via indexing in a kd-tree. Counts number of votes obtained by each model Types of matches: –Matches with correct model –Stray matches with incorrect model –Features that did not match well with any model Candidate matches
16
(x 1,y 1,s 1, 1 ) (x 2,y 2,s 2, 2 ) Transform predicted by this match: x = x 2 -x 1 y = y 2 -y 1 s = s 2 / s 1 = 2 - 1 Each match is represented by a dot in the space of 2D similarities (Hough space) xx yy ss Use of rich descriptors (SIFT) Lowe’99,’04
17
Prediction of position of model center after transform The space of transform parameters is discretized into ‘bins’ Coarse bins to limit boundary issues and have a low false- alarm rate for this stage We count the number of votes collected by each bin. Model: statistical estimation => learns fraction of matches in correct bin and neighbors Coarse Hough transform Model Test scene correct transformation
18
Correspondence or clutter ? RANSAC ‘RANdom SAmple Consensus’ – robust statistic for parameter estimation 2D affine transform : 6 parameters each sample contains 3 candidate correspondences. Model: –Probability p of detecting a feature –clutter: Poisson d d d Output of RANSAC : pose transformation + set of features correspondences
19
Consistency Consistency between observations and hypothesis foreground features ‘null’ assignments geometry appearance Consistency - appearanceConsistency - geometry
20
Learning foreground & background densities Ground truth pairs of matches are collected Gaussian densities, centered on the nomimal value that appearance / pose should have according to H Learning background densities is easy: match to random images. Moreels&Perona, IJCV, 2007
21
Score of an extended hypothesis Hypothesis: model + position observed features geometry + appearance database of models constant Consistency Prior on model and poses Assignments Votes per modelVotes per model pose bin Prior on assignments (before actual observations)
22
Experiments
23
UIUC database – Models 8 objects, 20 views/object
24
UIUC database – Test scenes 52 test scenes
25
Giuseppe Toys database – Models 61 objects, 1-2 views/object
26
Giuseppe Toys database – Test scenes 141 test scenes
27
a. Object found, correct pose Detection b. Object found, incorrect pose False alarm c. Wrong object found False alarm d. Object not found Non detection Performance evaluation Test image hand-labeled before the experiments
28
Results – UIUC database Test scenes Models from database ROC Threshold on probabilistic score.
29
Results – Giuseppe Toys database Lowe’99,’04 Lower false alarm rate - more systematic verification of geometry consistency - more consistent verification of geometric consistency undetected objects: features with poor Appearance distinctiveness Index to incorrect models - +
30
Results – Giuseppe Toys database
31
Coarse-to-fine strategy Probabilistic interpretation of each step Higher performance, comparable speed Conclusions
32
Result from ‘Google Images’ Looking for Grandma (or Brad Pitt) in collaboration with Alex Holub
33
Our result Looking for Grandma (or Brad Pitt)
34
Uses Viola-Jones face detection Face detection Confidence scores (M.Everingham) used to reject false alarms
35
Face warped to offset orientation and scale Patches extracted around face features Face representation Face representation by a long vector (eg. 19x11x11=2299) Dimensionality reduction to 50 dimensions via PCA
36
Similarity matrix in feature space
37
What face features are most useful ?
38
What size of patch is most useful ?
39
Distance optimization Goal: - small distance within images of same individual - larger distance between images of different person Optimization of following cost function Gradient descent used for optimization
40
Results – Tom Cruise
41
Results – Ronald Reagan
42
To conclude - Matlab demo
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.