A blog by Emmanuel Goossaert
During the Gold Rush of California in 1849, thousands of men were ready to put at stake all they had in the hope of a better life. The software industry that we are currently experiencing shows, to a certain extent, very similar patterns. Smartphone and source code have replaced the traditional pick and scoop, however the goals and hopes remained the same. But is there anything for developers to expect from this gold rush of software?
As part of my research on image segmentation, I have explored different methods for selecting areas in an image. Recently, I found a statistical color model based upon Lambertain surface reflectance. I have implemented this model using OpenCV 2.1. This article presents the results of some experiments I have run, along with my personal feelings about the model. At the end of the article, you will find links to the source code and to the research papers I used.
My implementation of the Active Appearance Models (AAMs) in C++ is almost done, it is called Paamela. I am currently fixing a couple design issues and finishing up the documentation. Even though I am still not sure whether or not I will make the code open source, I thought it would be nice to share what I have developed so far, in order to help other developers working on similar problems.
As part my C++ implementation of the AAMs (Active Appearance Models), I have been using OpenCV 2.1 full-time for about four weeks now. I thought I would share my feeling of the pros and cons about OpenCV. After reading this post, you should be able to know whether OpenCV is the solution you need or not.
In this article, I present a small raw socket daemon coded in C called “Knock-knock”, that allows port knocking to secure services under Linux. I have been using Knock-knock daily on a Ubuntu server for two years without any problem, and the code is small and simple enough for anybody to understand it.
Counting with MapReduce seems straightforward. All what is needed is to map the pairs to the same intermediate key, and leave the reduce take care of counting all the items. But wait, what if we have millions of items? Then one reducer, that is to say one process on one computer, will be forced to handle millions of pairs at once. Nonetheless this is going to be very slow, and all the interest of having a cluster will be missed, but there is something more important: what if the data is too big to fit in memory? Here I am showing how to count elements using MapReduce in a way that really split up the task between multiple workers.
Let us have a look at the solution discussed above. This solution counts items in a data in only one MapReduce iteration. Note that the values are replace with the value 1. Indeed, as counting does not require to keep track of the values, they are all changed to a common simple value to simplify computations. This solution seems pretty sweet, except that as we can see on Figure 1, reducing all the pairs to the same intermediate key gives one reducer, and one reducer only, a huge workload for counting the items. This can be efficient if the dataset is small. But there are cases in which the dataset is so big that it does not even fit into the memory of a single computer, or maybe it is so big that the computation on only one reducer is going to be very slow, and we need to know the count as soon as possible. As we will see in the next section, there is a way to improve workload balance along with computation time, at the cost of an additional iteration.
Simulated annealing is an optimization technique that finds an approximation of the global minimum of a function. When working on an optimization problem, a model and a cost function are designed specifically for this problem. By applying the simulated annealing technique to this cost function, an optimal solution can be found. In this article, I present the simulated annealing technique, I explain how it applies to the traveling salesman problem, and I perform experiments to understand how the different parameters control the details of the search for an optimal solution. I also provide an implementation in Python, along with graphic visualization of the solutions.
A few years back, thinking that you could have a cluster in your garage would have been crazy. Programming your own implementation of a reliable and powerful distributed system is feasible, but be ready to spend some months on it. Luckily, big companies and their need to handle increasing quantities of data led us to accessible solutions for cloud computing. The last groundbreaking solution in date, effective on clusters of cheap computers and developed by Google, is MapReduce. This article is yet another post on MapReduce, except that it is aimed at tech-savvy and non tech-savvy people, as it covers in details the different steps of a MapReduce iteration. It also explains how MapReduce is related to functional programming, why it enables parallel computing, and finally how the work is being distributed between workers during an iteration.
