incremental p can extend with Taylor series
a condition for corner
Copyright By PowCoder代写 加微信 powcoder
choose best feature/
image patch for tracking
image alignment
Mean-shift Tracking
move window throug whole image and calculate SSD/SAD tofind the most similar patch
#2 pyramid template matching
template matching
first scale the image, and then apply template matching to each scale
#3 model refinement
precondition have a good intial solution
fastest, locally optimal
slow, global solution
faster, locally optimal
problem definition p for translate
Lucas-Kanade
have a good intial guess p the guess p is very close to the final p
solve this
LK motion estimation is a special
case of LK image alignment
in KL alignment p is an affine transmation matrix
in LK motion estimation p is translate matrix
KLT Tracking
good feature
void smooth regions and edges
to make Hession matrix inversion the eigenvalues should be large
find corners staisfying two eigenvalues larger than threshold
the apply Lucas-Kanade method to each corner
update corner position
1. initial a position
2. compute mean-shift
3. update position
m(x) is a position of mean sample point g is guassian kernel, h is hyperparameter
for same spatial density add weights
shift: update:
target(q) and candidate(p(y)) descriptor normalized color histogram
measure the minimum distance/similarity(d(y)) = maximum cosine distance find y at that point
assuming a good inital and apply Taylor series expansion 整理之后,是mean-shift algorithm
classification
image classification
semantic segementation
instance sgemenation
object detection
Machine learning basic
supervised/unsupervised learning whether has labeled data
discrete/continuous for output
machin learning framework
given a training set of labeled examples{(xn,yn)},
estimate the prediction function f by minimizing
the prediction error on the training set
apply f to a never befor seen test example
x and output the prdiced value y=f(x)
identifying what is in the image
localizing and classigying one or more objects in an image
assigning a lable to every pixel in the iamge and treat
multiple objects of the same class as a single entity
binary label or multi-label classification
similar to semantic segementation but trading multiple
objects of the same class as distinct individual objects
definition
recognition
impage input feature representation tainable classifier ouput class label pipline
hand-crafted feature:
bag of feature
part-based models
motivation textue model
bags of words e.g. to understand rough topic about a book
1. extract local feature
2. learn “visual vocabulary”
3. quantize local features using visual vocabular
4. represent images in histogram by fequencies of “visual words”
extract descriptors from training set
clustering
trainable classifier
linear classifiers
simple to implement
decision boundaries not necessarily linear
works for any number of classes
non-parametric method
need good distance function
slow at test time
low-dimensional parametric representation
very fast at test time
cons works for two classes
K-NN select K nearest result and vote for the output
find a hyperplane that maximizes the marign between the positive and negative example
margin:2/||w||maximize margin = minize ||w||/2
maximize the margin when
classify training data correctly
for separable data
for non-separable data
maximize the margin and
minimize classification
mistakes(hinge loss function) in the result there will be some non-sperable data in 1/||w|| region
non-linear SVM map the data to higher dimensional feature space and data will be separable
non-linear SVM framwork is very powerful and flexible
find globally optimal solution
can work well with small traingin sample sizes
direct multi-class SVM achieve by combining two-class SVM instead of directly
computation, memory
Pedestrain detection
Face detection
Specify Object Model
Generate Hypotheses
Score Hypotheses
Resolve Detection
approaches
statistical template in bounding box
articulated parts model object is configuration of parts
hybrid template/parts model combine of the first two
find feature inthe box
3D-ish model
deformable 3D model
propse an alignment of the model to the image
find object parameters
approaches
sliding window
region-based proposal
mainly for gradient-based feature
rescore each propsed object based on the whole set
e.g. non-max suppressionapproaches
basic pipline
HOG&linear SVM
sytatistical template
gives score to each part and add them together
if the sum is ove a threshold, it will be a person
object model =
sum of scores of features
at fixed position
extract fixed-sized window at teach position and scale
compute HOG(histogram of gradient) feature within each window
score the window with a linear SVM classifier
perform non-maximua suppression to remove overlapping detections with lower scores resolve detection
score hypotheses
HOG feature descriptor
character for statis tical template approach
work well for non-deformable objects
fast detection
not so well for highly deformable objects
not robust to ocllusion
1. integral images for fast feature evalution
2. boosting for feature selection
3. attentional cascade for fast
non-face window rejection
以这个pixel向左上角话矩形,矩形内pixel的直累加到这个pixel的位置 fast to compute
最开始每个点的权重都相同
in each round, train one feature. Find the classifer
that achieves the lowest weight training error
for worry classified porints,
increasing their weight
after all the rounds, the final classifier as
linear combination of all wak classifier 精度与每个week classifier的权重成正比
early in the cascade, reject many negative examples, which not contain faces
chain classifiers are progressively more complex and have lower false positive rate
the detection rate and false positvie rate of the cascade 是每一级的乘积
H calls Hessian matrix
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com