Figure 1: WH, MSCR and RHSP on the single shot images.
In re-identification, the color is the most important feature. Here are some most frequently used color feature accumulation strategies. All of them can be applied in different color space (RGB,HSV,etc.).
Briefly speaking:
Multi-camera surveillance network illustration of Re-ID.[1]
(This outline is arranged from [2]. I’m still reading these papers and I’ll add some comment for several interesting methods in this week.)
The methods depend on the different data set. For example, methods related to the facial recognition is not suitable for the Market1501.
Since the spectrum auction of Federal Communications Commission (FCC) in 1989, combinatorial auction becomes a very active field between economy and computer science [1]. On the other hand, the posted price mechanism is the cornerstone of the E-commerce, especially in online shopping. Therefore, it is meaningful to design combinatorial auction mechanisms via posted price.
In combinatorial auction, the valuation of bundle is not always equal to the sum of items in bundle naively (e.g. subaddictive, submodular, XOS, etc.). The aim of combinatorial auction is to maximize the social welfare by allocating the items. It is obvious that we can solve this problem by running VCG algorithm. However, VCG algorithm is computationally infeasible for complex problems [2]. Thus, a natural idea appears: can we design an approximate algorithm which can be implemented in polynomial time?
These days I tried to install the TensorFlow with VirtualEnv and encountered several problems caused by Anaconda. Fortunately I found the solution.
If your computer installed Anaconda before you install the virtualenv, then you should use the1
conda install virtualenv
instead of the
1 | pip install virtualenv |
Then everything goes well now :D
Illustrated as the banner of the game Atletico vs Leicester City
Match Thread (in Chinese)
要做的事堆积如山，这样的感觉真是久违了呢（笑）
第一回合已经顺利拿下，希望第二回合继续稳扎稳打
Our video presentation of Recommender Systems - Beyond Matrix Completion
Video: Shuyao Chen
Audio: Aidon Blong
Script: Aidon Blong, Shuyao Chen, Felix Udanyi
Recently I’m reading some papers related to the sequential probability ratio test(SPRT). Here is some notes about the concepts from statistical hypothesis testing.
I don’t have any background knowledge about statistics. If you find any mistake in this notes please tell me :)
Define the distortion function as:
$$ J(c,u)=\sum{i=1}^{m}||x^{(i)}-u{c^(i)}||^{2} $$
Function J is the square sum of the distance between centroid and corresponding vectors. The aim of K-means algorithm is to minimize the J. It’s easy to find that there are 2 methods to minimize the J:
When J is near to 0, which means the sum of the distance between the centroid and vectors is decreasing. Thus, we can set a tolerance variable to judge the similarity of 2 centroids. When the distance between 2 mean centroids is less or equal to the tolerance, it means the clustering consequence is converged. Depend on the sklearn library, we set the tolerence = 0.0001
Although the function J is not a convex function, which means we may converge to a local optimal solution. But J shows that the ordinary K-means algorithm can converge to a minimal value.
