training image from google, testing on Caltech dataset
unsupervised learning algorithm
pLSA
no spatial information
ABS-pLSA
latent variable x (grid centroid coordinate)
grid methodology, still translation and scaling variant
TSI-pLSA
latent variable c:(centroid, xscale, yscale)
1. apply to pLSA
2. fit a mixture of Gaussian with k=(1, 2, ...K) components to the location of the region, weighted by p(w|z)
3.
Posted in
L10,
lecture notes
A. Abstract
Modeling and recognizing landmarks at world-scale is a useful yet challenging task. There exists no readily available list of worldwide landmarks. Obtaining reliable visual models for each landmark can also pose problems, and efficiency is another challenge for such a large scale system.
This paper leverages the vast amount of multimedia data on the web, the availability of an Internet image search engine, and advances in object recognition and clustering techniques, to address these issues.
First, a comprehensive list of landmarks is mined from two sources: (1) ∼ 20million GPS-tagged photos and (2) online tour guide web pages. Candidate images for each landmark are then obtained from photo sharing websites or by querying an image search engine.
Second, landmark visual models are built by pruning candidate images using efficient image matching and unsupervised clustering techniques.
Finally, the landmarks and their visual models are validated by checking authorship of their member images. The resulting landmark recognition engine incorporates 5312 landmarks from 1259 cities in 144 countries.
The experiments demonstrate that the engine can deliver satisfactory recognition performance with high efficiency.
B. Notes
Hierarchical Clustering
application
Google Landmark, upload photos to search landmark names, positions, etc.
Posted in
L7,
lecture notes,
must read
Image segmentation
- spatial space
- region growing
- watershed
- feature space
- kmeans
- graph based
Min Cuts
NP-complete -> eigen problem to approximate
recursively two-way Ncut
- when eigen value still small
- when inter-group pixels are not very similar (still can be segmented)
simultaneous K-way cut
- iterative merge
Posted in
L6,
lecture notes,
must read
LLE
constrains to avoid many sols.
1. mean = 0
2. Y cross YT = I
IsoMap
1. still use k-nn
2.
compute pairwise geodesic distance
3. MDS
Posted in
L5,
lecture notes
Fisher face (FLD) is different from PCA (principle component analysis), it is supervised.
So that data points will be projected to low dimensional space which can separate different type of data.
Posted in
L4,
lecture notes
A. Summary (mostly come from from ABSTRACT)
This paper describe latent Dirichlet allocation (LDA), a generative probabilistic model for collections of discrete data such as text corpora. LDA is a three-level hierarchical Bayesian model, in which each item of a collection is modeled as a finite mixture over an underlying set of topics.
Each topic is, in turn, modeled as an infinite mixture over an underlying set of topic probabilities. In the context of text modeling, the topic probabilities provide an explicit representation of a document.
We present efficient approximate inference techniques based on variational methods and an EM algorithm for empirical Bayes parameter estimation. We report results in document modeling, text classification, and collaborative filtering, comparing to a mixture of unigrams model and the probabilistic LSI model.
B. Note
part1: corpus generative process model
LDA assumes the following generative process for each document
w in a corpus D:
1. Choose N ~ Poisson(ξ).
2. Choose θ ~ Dir(α).
3. For each of the N words w
n:
(a) Choose a topic z
n ~ Multinomial(θ).
(b) Choose a word w
n from p(w
n | z
n,β), a multinomial probability conditioned on the topic z
n.
part2: Estimate parameter
Maximizing p(D | α,β) instead of maximizing p(α,β | D), because they are equal (by Bayesian theorem)
but p(
w | α,β) is intractable, so they used a variational distribution to approximate γ,φ
Posted in
L3,
lecture notes,
must read
A. Summary (mostly come from from ABSTRACT)
Probabilistic Latent Semantic Indexing is a novel approach to automated document indexing which is based on a statistical latent class model for factor analysis of count data.
Fitted from a training corpus of text do cuments by a generalization of the Expectation Maximization algorithm, the utilized model is able to deal with domain-specific synonymy
as well as with polysemous words.
In contrast to standard Latent Semantic Indexing (LSI) by Singular Value Decomposition, the probabilistic variant has a solid statistical foundation and defines a proper generative data model.
Retrieval experiments on a number of test collections indicate substantial performance gains over direct term matching methods as well as over LSI.
In particular, the combination of models with different dimensionalities has proven to be advantageous.
B. Note
LSA
document->concept->word
SVD(singular value secomposition)
pLSA
EM algorithm
Posted in
L3,
lecture notes,
must read
what is the most important characteristic of a feature
efficiency? yes
dimension length? (yes but usually long for image and video)
invariant? !! scale, rotation
*local feature began widely used about when '01
bag of feature points
problem: efficiency, thousands of them need to be matched
Local Feature = Detector + Descriptor (where are them, how to represent)
Detector:
point-based
region-based
scale invariant: LoG, DoG(for low quality), Harris-Laplace
affine invariant: Harris-Affine, Hessian-Affine, ..., etc
*DoG(Difference of Gaussian) detector:
apply various kernels and scales
deal with scale-invariant
Descriptor:
*SIFT: empirically 4x4(windows)x8(bins)=128 dimensions
deal with rotation-invariant
Visual Word
quantized local descriptor (by clustering) e.g. SIFT
*some issues
the more, the better?
yes, but saturates in object retrieval, or even degrade in classification
efficiency issue.
Algorithm-wise or Parallelize
sampling method
sparse, dense, or random
feature selection and reduction
pLSA and LDA
low recalls
soft-assignment, query expansion
Posted in
L2,
lecture notes
