hoggles: visualizing object detection features (to be appeared in iccv 2013) carl vondrick aditya...
TRANSCRIPT
HOGgles: Visualizing Object Detection Features (to be appeared in ICCV 2013)Carl Vondrick Aditya Khosla Tomasz Malisiewicz and Antonio Torralba ,MIT
Presented By: Yonatan Dishon Nov 2013
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
So why HOGgles?
Image from: C. Vondrick, A. Khosla, T. Malisiewicz, A. Torralba. "HOGgles: Visualizing Object Detection Features" 2013
So why HOGgles?
!?Maybe I should put my HOGgles!!HOGgles Visualization
So why HOGgles? This is a visualization of the descriptor space – this is
what a classification/detection algorithm sees! So which of the following would you classify as a car?
Remember Humans are the “perfect” classifier!
Object categories (PASCAL VOC): Airplane, Bicycle, Bird, Boat , Bottle, Bus, Car, Cat, Chair, Cow, Table, Dog, Horse, Motorbike, Person, Potted Plant, Sheep, Sofa, Train, TV/Monitor
1 2 3 4 5 6 7
Motivation So why did my detector failed?
Training set Maybe not a good one? Learning Algorithm Maybe should have been
different? Features Maybe there are better features for this
kind of problem?
Visualizing the Feature space can bring us to an intuitive understanding of our detection system limitations and failures
HOGgles Contributions A tool to explain some of the failure of object
detection systems. Algorithm to present the feature space of
object detectors (general features – not only HOG!).
4 different algorithms to do so are presented. Public feature visualization toolbox -
http://web.mit.edu/vondrick/ihog/#code
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
Related Work (1) Reconstruct an image given keypoint of SIFT
based on a huge database
P. Weinzaepfel, H. J´egou, and P. P´erez. Reconstructing an image from its local descriptors. In CVPR, 2011
Image from: P. Weinzaepfel, H. J´egou, and P. P´erez. Reconstructing an image from its local descriptors. In CVPR, 2011.
Related Work (1)
Related Work (1)
original imagecopied patches blending interpolation
Reconstructing an image from its local descriptors, Philippe Weinzaepfel, Hervé Jégou and Patrick Pérez, Proc. IEEE CVPR’11.
calculate SIFT elliptic region of interest
affine normalization to square patch
Slide credit: Ezgi Mercan
Related Work (1)
Image from: P. Weinzaepfel, H. J´egou, and P. P´erez. Reconstructing an image from its local descriptors. In CVPR, 2011.
Website for more examples : http://www.irisa.fr/texmex/people/jegou/projects/reconstructing/index.html
Related Work (2) Reconstruct an image given only LBP features
E. d’Angelo, A. Alahi, and P. Vandergheynst. Beyond bits: Reconstructing images from local binary descriptors. ICPR, 2012. 2
Using LBD descriptors (Local Binary Descriptors) – BRIEF and FREAK
No external information! Reconstruction as a regularized inverse problem
A. Alahi, R. Ortiz, and P. Vandergheynst. FREAK: Fast Retina Keypoint. In IEEE Conference on Computer Vision and Pattern Recognition (To Appear), 2012.
M. Calonder, V. Lepetit, C. Strecha, and P. Fua. BRIEF: Binary Robust Independent Elementary Features. Computer Vision–ECCV 2010, pages 778–792, 2010.
Related Work (2)
Image reconstruction results of FREAK (top row) and BRIEF (bottom row) descriptors.
E. d’Angelo, A. Alahi, and P. Vandergheynst. Beyond bits: Reconstructing images from local binary descriptors. ICPR, 2012. 2
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
HOGgles – under the hood The problem of feature visualization as a
feature inversion problem.Given a feature vector - what was the
image/patch that created it? Let be an image and be the
corresponding HOG feature descriptor. is a many to one function.
The inversion problem cannot be solved analytically!
Dx R y x
HOGgles under the hood The problem is formalized as an optimization
problem – given a descriptor y we seek an image x that minimize:
This function isn’t convex Trying to find a minima with ordinary
optimization algorithms didn’t work (Steepest decent and Newton’s method).
2* 1
2argmin
Dx R
x y x y
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
HOGgles under the hood 4 algorithms are showed – 3 as a baselines
and one is offered as the main algorithm. Baseline 1: Exampler LDA
Baseline 2: Ridge Regression
Baseline 3: Direct Optimization
Main Algorithm : Paired Dictionary
HOGgles under the hood (Baseline 1 -ELDA)
Baseline 1: Exampler LDA (B.Hariharan, , J.Malik & D.Ramanan ECCV 2012)
HOG inverse is the average of the top K detections of the ELDA detector in RGB space .
1 y
HOGgles under the hood (Baseline 1 -ELDA)
Slide credit: Ezgi Mercan
HOGgles under the hood (Baseline 1 -ELDA) PROs:
Simple. Surprisingly accurate results! Even when the
database doesn’t contain the category of HOG template!
CONs: Computationally expensive! – running an object
detector over a large database Yields blurred results.
HOGgles under the hood (Baseline 2 –Ridge Regression) Baseline 2: Ridge Regression
Statistically most likely image given a HOG descriptor.
Calculating the most probable grayscale image given its HOG feature.
Modeling as The HOG inverse (the visualization) is given by
are estimated on a large database Single matrix multiplication!
,N X | YP
1 1B XY YY Y Xy y
Xand
X,YP
HOGgles under the hood (Baseline 2 –Ridge Regression) PROs:
Simple – this is a matrix multiplication! Very fast! – under a second for inversion.
CONs: Inversion yields blur images.
HOGgles under the hood (Baseline 3 -Direct) Baseline 3: Direct Optimization
Describing a natural image basis Any image can be encoded by coefficients in this basis: And we wish to find
D KU R Dx R KR
x U
2*
2argminKR
U y
HOGgles under the hood (Baseline 3 -Direct) PROs:
Recover high frequencies CONs:
adding noise to the image
HOGgles under the hood (the main algorithm –Pair Dictionary) Let be an image be its HOG
descriptor. Suppose we write x and y in terms of bases and respectively where are shared coefficients
y can be projected to V basis and then to image basis U
Dx R dand y RD KU R
d KV R
HOGgles under the hood (Main Algorithm –PairDict)
d KV R
D KU R
HOGgles under the hood (Baseline 4 –PairDict) How the bases U and V are found?
Solving Pair dictionaries learning problem.
The objective is simplified to a standard sparse coding and dictionary learning problem, optimized with SPAMS
J. Mairal, F. Bach, J. Ponce, and G. Sapiro. Online dictionary learning for sparse coding. In ICML, 2009. 4SPAMS – SParse Modeling Software, Code available at http://spams-devel.gforge.inria.fr/
*
*
HOGgles under the hood (Baseline 4 –PairDict) Dictionaries optimization time is a few hours (offline). U and V are estimated with dictionary size Training samples from a large database
examples for U & V pairs – notice the correlation of the dictionaries
310K 610N
HOGgles visual results
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
HOGgles – Limitations There exist a better visualization than paired
dictionaries although it may not be traceable to construct it.
On Recursive iterative solution some high frequencies are lost.
HOGgles – Limitations – cont. Inversion is sensitive to the HOG
dimensionality
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
Evaluation of Inversions PASCAL VOC 2011 dataset. Inverting patches correspond to objects. Quantitative evaluation:
How well each pixel in x is reconstructed from y by each Algorithm?
Qualitative evaluation: How well the high level content is saved?
(Human research using MTurk platform).
Evaluation of InversionsQuantitative Mean normalized cross correlation of inverse
image to ground truth. (Higher is better , max. is 1)
bicycle bottle car cat chair table motorbike person Mean0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.56
0.67
0.64
0.71
0.62 0.610.59
0.65 0.64
Evaluation of Performance
ELDA Ridge Direct PairDict
Evaluation of InversionsQualitative MTurk workers where asked to classified
inversions to 1 of 20 categories. MIT PhD. In CV refers as experts. Trying the
same task with HOG Glyphs.
(*) Numbers are percentage classified correctly, chance is 0.05
Evaluation of InversionsQualitative
Graphs credit : Ezgi Mercan
bicycle bottle car cat chair table motorbike person Mean
-0.10
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.31
0.45
0.59
0.20
0.39
0.24 0.22
0.68
0.38
Evaluation of Vizualization PerformanceELDA Ridge Direct PairDict Glyph Expert
Evaluation of InversionsQualitative Gliphs vs. HOGgles
Today talk Motivation Related Work HOGgles under the hood.
3 Baselines + main algorithm Limitations Quantative + Qualitative evaluation
Human+HOG detector Paper Conclusion Future Development
Human+HOG detector Get Insight on performance of HOG with the
perfect learning algorithm (people).
Large Human Expirement consist of: MTurk workers Dataset – Top detections from DPM on PASCAL
2007 VOC. 5000 windows per category. 20% are true positive. 25 votes per window.
Human+HOG detectorResults
Paper Conclusions DPM is operating very close to the
performance limit of HOG. HOG may be too lossy of a descriptor for high
performance object detection. The features we are using are the one to
blame in current novel object detection algorithms.
To advance to the next level recognition – finer details and higher level information capturing features are needed to be built.
Written Level of the Paper Pros:
Well written Well referenced Novel solution Large and detailed Human experiment Website.
Cons: Details/examples on baseline algorithms are
lacking The chosen algorithm settings are lacking Some conclusions on the Human+HOG are
questionable
Future development Applying visualization on other descriptors. Color reconstruction of HOG features. Developing new feature that can be more
discriminate. Developing Algorithms that better model the
interaction of the simple atoms of the image.
THANK YOU!QUESTIONS?
Links PASCAL 2 challenge Deformable Part Model HOGgles