Tracking

[trackingsparse-blog]

[trackingsparse-blog]

LaRank vs OLaRank

Synthetic data with dimensions $[250,2]$.

Results (train/test):

LaRank  .992/.964

OLaRank online manner: 0.552, After learning 0.964/0.956.

[trackingsparse-blog]

ibogun/Multiclass SVM

Winner-takes-it-all multiclass svm using structured output SVM. - Gist is a simple way to share snippets of text and code with others.

[trackingsparse-blog]

Loss functions

Hinge loss $$\max \{0,1-y_n w^T x_n \}$$

Log loss $$\log [1+ \exp{(-y_n w^T x_n) }]$$

Exponential loss $$\exp{(-y_n w^T x_n) }$$

 Object localization $$ \Delta(y,y')=1- \frac{|y \cap y'|}{|y \cup y'|}$$

[trackingsparse-blog]

Rotational invariance

[trackingsparse-blog]

Rotational non-invariance

[trackingsparse-blog]

Exploratory analysis for COSFIRE operators vs sparse portraits.

[trackingsparse-blog]

Exploratory analysis of COSFIRE filters on sparse portraits

Problem: find a way how COSFIRE filters could detect the object in a robust way?

what does define a good filter?

How much tuples are necessary for operator to be useful?

Steps used to perform analysis:

Calculate COSFIRE operator at each location in the middle of each ground truth data point.

Visualize it and compare to the image itself.

Save resulting images to create a video

[trackingsparse-blog]

sparse velocity (speed decomposition)

[trackingsparse-blog]

Video c09

[trackingsparse-blog]

Video c01

[trackingsparse-blog]

c15 prior normalization

[trackingsparse-blog]

c09 prior normalization

[trackingsparse-blog]

c01 prior normalization

[trackingsparse-blog]

Sparse portraits in different color spaces part 2.

[trackingsparse-blog]

Sparse portraits in different color spaces part 1.

[trackingsparse-blog]

Difference between video decomposition if sharpening is used. This is video 1, frames 1,20,30,35.