Computer Vision Group

Computer Vision GroupUniversity of California Berkeley

Recognizing objects and actions in images and video

Jitendra Malik

U.C. Berkeley


Collaborators

• Grouping: Jianbo Shi, Serge Belongie, Thomas Leung

• Ecological Statistics: Charless Fowlkes, David Martin, Xiaofeng Ren

• Recognition: Serge Belongie, Jan Puzicha, Greg Mori


Motivation

• Interaction with multimedia needs to be in terms that are significant to humans: objects, actions, events..

• Feature based CBIR systems (e.g. color histograms) are a failure for this purpose

• Presently, the best solutions use text / speech recognition/ human annotation

• Goal: Auto-annotation


From images/video to objects

Labeled sets: tiger, grass etc




ConsistencyA

B C

• A,C are refinements of B• A,C are mutual refinements • A,B,C represent the same percept

• Attention accounts for differences

Image

BG L-bird R-bird

grass bush

headeye

beakfar body

headeye

beak body

Perceptual organization forms a tree:

Two segmentations are consistent when they can beexplained by the samesegmentation tree (i.e. theycould be derived from a single perceptual organization).


What enables us to parse a scene?

– Low level cues• Color/texture• Contours• Motion

– Mid level cues• T-junctions• Convexity

– High level Cues• Familiar Object• Familiar Motion


Outline

• Finding boundaries

• Recognizing objects

• Recognizing actions


Finding boundaries: Is texture a problem or a solution?

image orientation energy


Statistically optimal contour detection

• Use humans to segment a large collection of natural images.

• Train a classifier for the contour/non-contour classification using orientation energy and texture gradient as features.


Orientation Energy

• Gaussian 2nd derivative and its Hilbert pair

•

• Can detect combination of bar and edge features; also insensitive to linear shading [Perona&Malik 90]

• Multiple scales

22 )()( evenodd fIfIOE


Texture gradient = Chi square distance between texton histograms in half disks across edge

i

j

k

K

m ji

jiji mhmh

mhmhhh

1

22

)()()]()([

21),(Chi-square

0.1

0.8





ROC curve for local boundary detection


Outline





Biological Shape

• D’Arcy Thompson: On Growth and Form, 1917– studied transformations between shapes of organisms


Deformable Templates: Related Work

• Fischler & Elschlager (1973)

• Grenander et al. (1991)

• von der Malsburg (1993)


Matching Framework

• Find correspondences between points on shape

• Fast pruning

• Estimate transformation & measure similarity

model target

...


Comparing Pointsets


Shape ContextCount the number of points inside each bin, e.g.:

Count = 4

Count = 10

...

Compact representation of distribution of points relative to each point


Shape Context


Shape Contexts

• Invariant under translation and scale

• Can be made invariant to rotation by using local tangent orientation frame

• Tolerant to small affine distortion– Log-polar bins make spatial blur proportional to r

Cf. Spin Images (Johnson & Hebert) - range image registration


Comparing Shape ContextsCompute matching costs using Chi Squared distance:

Recover correspondences by solving linear assignment problem with costs Cij

[Jonker & Volgenant 1987]


Matching Framework


• Fast pruning


model target

...


Fast pruning

• Find best match for the shape context at only a few random points and add up cost

),(minarg

),(),(

2*

*

1

2

ui

jqueryui

ij

query

r

jiquery

SCSCSC

SCSCSSdist


Snodgrass Results


Results


Matching Framework


• Fast pruning


model target

...


• 2D counterpart to cubic spline:

• Minimizes bending energy:

• Solve by inverting linear system

• Can be regularized when data is inexact

Thin Plate Spline Model

Duchon (1977), Meinguet (1979), Wahba (1991)


MatchingExample

model target


Outlier Test Example


Synthetic Test Results

Fish - deformation + noise Fish - deformation + outliers

ICP Shape Context Chui & Rangarajan


Terms in Similarity Score• Shape Context difference

• Local Image appearance difference– orientation– gray-level correlation in Gaussian window– … (many more possible)

• Bending energy


Object Recognition Experiments

• Handwritten digits

• COIL 3D objects (Nayar-Murase)

• Human body configurations

• Trademarks


Handwritten Digit Recognition

• MNIST 60 000: – linear: 12.0%– 40 PCA+ quad: 3.3%– 1000 RBF +linear: 3.6%– K-NN: 5%– K-NN (deskewed): 2.4%– K-NN (tangent dist.): 1.1%– SVM: 1.1%– LeNet 5: 0.95%

• MNIST 600 000 (distortions): – LeNet 5: 0.8%– SVM: 0.8%– Boosted LeNet 4: 0.7%

• MNIST 20 000: – K-NN, Shape Context

matching: 0.63%



Results: Digit Recognition

1-NN classifier using:Shape context + 0.3 * bending + 1.6 * image appearance


COIL Object Database


Error vs. Number of Views


Prototypes Selected for 2 Categories

Details in Belongie, Malik & Puzicha (NIPS2000)


Editing: K-medoids

• Input: similarity matrix

• Select: K prototypes

• Minimize: mean distance to nearest prototype

• Algorithm: – iterative– split cluster with most errors

• Result: Adaptive distribution of resources (cfr. aspect graphs)


Error vs. Number of Views


Human body configurations


Deformable Matching

• Kinematic chain-based deformation model

• Use iterations of correspondence and deformation

• Keypoints on exemplars are deformed to locations on query image


Results


Trademark Similarity


Recognizing objects in scenes


Shape matching using multi-scale scanning

• Shape context computation (10 Mops)– Scales * key-points * contour-points (10*100*10000)

• Multi scale coarse matching (100 Gops)– Scales * objects * views * samples * key-points* dim-sc

(10*1000*10*100*100*100)

• Deform into alignment (1 Gops)– Image-objects * shortlist * (samples)^2 *dim-sc

(10*100*10000*100)


Shape matching using grouping

• Complexity determining step: find approx. nearest neighbors of 10^2 query points in a set of 10^6 stored points in the 100 dimensional space of shape contexts.

• Naïve bound of 10^9 can be much improved using ideas from theoretical CS (Johnson-Lindenstrauss, Indyk-Motwani etc)


Putting grouping/segmentation on a sound foundation

• Construct a dataset of human segmented images

• Measure the conditional probability distribution of various Gestalt grouping factors

• Incorporate these in an inference algorithm


Outline





Examples of Actions• Movement and posture change

– run, walk, crawl, jump, hop, swim, skate, sit, stand, kneel, lie, dance (various), …

• Object manipulation– pick, carry, hold, lift, throw, catch, push, pull, write, type, touch, hit,

press, stroke, shake, stir, turn, eat, drink, cut, stab, kick, point, drive, bike, insert, extract, juggle, play musical instrument (various)…

• Conversational gesture– point, …

• Sign Language


Activities and Situation Classification

• Example: Withdrawing money from an ATM

• Activities constructed by composing actions. Partial order plans may be a good model.

• Activities may involve multiple agents

• Detecting unusual situations or activity patterns is facilitated by the video activity transform


On the difficulty of action detection

• Number of categories

• Intra-Category variation– Clothing– Illumination– Person performing the action

• Inter-Category variation


Objects in space Actions in spacetime

• Segment/Region-of-interest

• Features (points, curves, wavelet coefficients..)

• Correspondence and deform into alignment

• Recover parameters of generative model

• Discriminative classifier

• Segment/volume-of-interest

• Features (points, curves, wavelets, motion vectors..)

• Correspondence and deform into alignment

• Recover parameters of generative model

• Discriminative classifier


Key cues for action recognition

• “Morpho-kinesics” of action (shape and movement of the body)

• Identity of the object/s

• Activity context


Image/Video Stick figure Action

• Stick figures can be specified in a variety of ways or at various resolutions (deg of freedom)– 2D joint positions– 3D joint positions– Joint angles

• Complete representation

• Evidence that it is effectively computable


Tracking by Repeated Finding


Achievable goals in 3 years

• Reasonable competence at object recognition at crude category level (~1000)

• Detection/Tracking of humans as kinematic chains, assuming adequate resolution.

• Recognition of ~10-100 actions and compositions thereof.

Documents

Computer Vision Group