cvpr/iccv 09 paper reading dan wang nov. 6, 2009
TRANSCRIPT
Papers
• CVPR:– Learning Color and Locality Cues for Moving
Object Detection and Segmentation
• ICCV: – Texel-based Texture Segmentation
Learning Color and Locality Cues for Moving Object Detection and Segmentation
Feng Liu and Michael Gleicher
Author 1 Feng Liu
• Computer Sciences Department at the University of Wisconsin, Madison
• Graduate student• Publications
– Feng Liu, Michael Gleicher, Hailin Jin and Aseem Agarwala. Content-Preserving Warps for 3D Video Stabilization. ACM SIGGRAPH 2009
– Feng Liu, Yu-hen Hu and Michael Gleicher. Discovering Panoramas in Web Videos. ACM Multimedia 2008
– Feng Liu and Michael Gleicher. Texture-Consistent Shadow Removal. ECCV 2008
– Feng Liu and Michael Gleicher. Video Retargeting: Automating Pan and Scan. ACM Multimedia 2006
Author 2 Michael Gleicher
• Computer Sciences Department at the University of Wisconsin, Madison
• Professor• Positions – 2009 – present – Professor– 2004 – 2009 – Associate Professor– 1998 - 2004 – Assistant Professor
• http://pages.cs.wisc.edu/~gleicher/CV.pdf
Problem of Previous Methods
• Most previous automatic methods rely on methods rely on object or camera motionobject or camera motion to detect the moving object.
• Small motion Small motion of object or camera do not provide sufficient information for these methods.
Abstract• The paper presents an algorithm for automatically
detecting and segmenting a moving object from a monocular video.
• Existing method: Rely on motion to detect moving object. When motion is sparse and insufficient, it fails…
• What does this paper do ?– Unsupervised algorithm to learn object color and locality
cues from the sparse motion information.• How to do?– Detect key frames, sub-objects– Learning from the sub-objects: color and locality cues– Combining cues in MRF framework
• What is “Moving ObjectMoving Object”?– Some compact regions with different apparent
motion from background.
• How to detect moving object– Estimate the global motion in a video– Calculate the discrepancy at each between the
object motion and global motion
Detect Moving Object
• Model: – Use Homography to model the global motion
between two consecutive frames.
• Feature:– Use a SIFT feature-based method to estimate the
homography
projective transformationDetails in [19]
Key Frame Extraction
• Definition– The frame where a moving object or its part can
be reliably inferred from motion cues.– Motion cues are likely reliable when they are
strong and compact.
Segment Moving Sub-objects from Key Frames
• Problem– Not all pixels of the moving object have significant
motion cues. Motion cues are Sparse!
Segment Moving Sub-objects from Key Frames
• Solution– Neighboring pixels are likely to have the same
label.– Neighboring pixels with similar colors are more
likely to have the same label.
• Solved by Graph cut algorithm
Segment Moving Sub-objects from Key Frames
• MRF priors on labels to model the interaction
Pixel label
Number of pixels Neighbor
of pixel i
Segment Moving Sub-objects from Key Frames
• The likelihood of image I given a labeling can be modeled as follows:
Learning Color and Locality Cues
• Assumption– The moving sub-objects from all the key frames
form a complete sampling of the moving objects.
• Procedure– Lab color space– Build GMM
Learning Color and Locality Cues
• The spatial affinity of pixel I to the moving object:
• Location likelihood:
F: sub-objectparameter
position
Discussion• Extracting moving object from a video with less object
and camera motion is not easier than more objects and camera motion
• Contribution– Unsupervised?
• Currently– Off-line– Motion estimation is time consuming
• Future– Parameters– Background modeling
Authors
• Sinisa Todorovic– Assistant Professor– School of EECS, Oregon State University – Publication: CVPR 2009, ICCV 2009, TPAMI 2008...
• Narendra Ahuja– Donald Biggar Willet Professor– Beckman Institute , UIUC– Publication: ICCV 2009, IJCV 2008, TPAMI…
Problem
• Given an arbitrary image, segment all texture subimages– Texture = Spatial repetition of texture elements, i.e.,texels– Texels are not identical, but only statistically similar to one
another.– Texel placement along the texture surface is not periodic, but
only statistically uniform.– Texels in the image are not homogeneous, but regions that
may contain subregions.
Rational
• Texels occupy image regions– If the image contains texture, many regions will
have similar properties• color, shape, layout of subregions• orientation, relative displacements
– The pdf of region properties will have modes
Texture detection and segmentation
Detection of modes of pdf of region properties
Contributions
• No assumptions about the pdf of texel properties
• Both appearance and placement of the texels are allowed to be stochastic and correlated
• New hierarchical, adaptive-bandwidth kernel to capture texel structural properties
Method Description
• Define a feature space of region propertiesregion properties• Descriptor of each region = Data point in the
feature space• Partition the feature space Partition the feature space into bins by Voronoi
tessellation• Run the meanshift with the new, hierarchical new, hierarchical
kernelkernel• Regions under a pdf mode comprise the texture
subimage
The feature space of region properties
• Image hierarchical structure
• Use multiscale segmentation algorithm[1], [4]
The feature space of region properties
• A descriptor vector of prperties xi of image region i1. Average contrast across i’s boundary2. Area3. Standard deviation of i’s children area4. Displacement vector between the centroids of i and its
parent region5. Perimeter6. Aspect ratio of intercepts of i’s principal axes with the
i’s boundary7. Orientation: the angle between principal axes and x-axis8. Centroid of I
PCA 95% Not scale and rotation invariance
Voronoi Diagram
• Definition:– Let P = {p1,p2, ... , pn} be a set of points in the plane (or in
any dimensional space), which we call sites. – Define V(pi), the Voronoi cell for pi, to be the set of points
q in the plane that are closer to pi than to any other site. That is, the Voronoi cell for pi is defined to be:
• V(pi ) = {q | dist(pi,q) < dist(pj, q), for j != i}:
• Anyway, it can partition the feature space…Anyway, it can partition the feature space…• http://www.dma.fi.upm.es/mabellanas/tfcs/fvd/voro
noi.html
Gaussian kernel
Hierarchical kernel
• Motivation– Texels, in general, are not homogenous-intensity
regions, but may contain hierarchically embedded subregions.
– Since region descriptors represent image regions, we can define hierarchical relationships between the descriptors based on the embedding of corresponding smaller regions within larger regions in the image.
• Voronoi partitioning of the feature space– Suppose x belongs to Bi, and x’ belongs to Bj
•Arbitrary points
•Region descriptors
x x’
xi xj
• x belongs to Bi . Compute by finding the maximum subtree isomorphism between two trees rooted at at xi and xj as:
Experimental Evaluation
• Qualitative Evaluation• Quantitative evaluation– G: the area of true texture– D: the area of a subimage that our approach
segments– Segmentation error per texture:
• (1) 100 collages of randomly mosaicked, 111 distinct Brodatz textures, where each texture occupies at least 1/6 of the collage
• (2) 180 collages of randomly mosaicked, 140 distinct Prague textures from 10 thematic classes (e.g., flowers, plants, rocks, textile, wood, etc.), where each texture occupies at least 1/6 of the collage .
• (3) 100 Aerial-Produce images, where 50 aerial images show housing developments, agricultural fields, and landscapes, and 50 images show produce aisles in supermarkets .
Quantitative evaluation(Brodaz)
93.3%+3.7
structural properties vs ignoring them 77.9%+4.1 Simple Gaussian kernel, Variable bandwidth
Variable bandwidth in this paper 62.3%+7.8 Simple Gaussian kernel, Variable bandwidth [7]
• Quantitative evaluation– G: the area of true texture– D: the area of a subimage that our approach
segments– Segmentation error per texture:
Quantitative evaluation(Prague)– G: the area of true texture– D: the area of a subimage that our approach segments– CS: Correct segmentation– OS: Over segmentation means G that is split into smaller regions D– US: DG– ME counts every G that does not belong to CS– NE counts every D that does not belong to US
“The state-of-the-art unsrpervised texture
segmentation”
Conclusion• We have presented a texel-based approach to
segmenting image parts occupied by distinct textures. This is done by capturing intrinsic and intrinsic and placement placement properties of distinct groups of texels.
• Experimental evaluation on texture mosaics and real-world images suggests that capturing structural properties structural properties of texels is very important for texture segmentation.
• To account for texel substructure, we have derived and used a hierarchical, variable a hierarchical, variable bandwidth kernel bandwidth kernel in the meanshift.
Slices are partly from Sinisa Todorovic