blog posts

We often confuse our tools for our craft. Tools help you practice your craft, but they do not make you a good craftsman. A good craftsman has many different tools in her pocket, and she judiciously uses the one that is appropriate for the job

I am often amused by wars over programming languages. People have very strong opinions about which one is better. The truth is that a programmer should chose the language appropriate for their task. Want to rapidly build a web app ? Try Ruby on Rails or Django. Want to write high performance code for an embedded device. Try C for Computer Vision.

In computer vision, we are faced with similar choices. Which tool should a computer vision engineer / programmer learn — OpenCV using C++, OpenCV using Python, or MATLAB ? The good news is that today we have a few options to choose from! A decade back there were no good libraries for computer vision. If you wanted to learn, you just picked up a book and started coding your own mini library of computer vision algorithms. Thankfully, things are much better now.

If you are a beginner, I would suggest using the path of least resistance and picking a tool you are familiar with. If you are a python programmer, use OpenCV with Python. If you know C++, use C++ with OpenCV. The same holds true for MATLAB. That said in a few months you will no longer be a beginner. You may be looking to apply this newly acquired knowledge to the real world. You could be thinking of a new side project, or looking for a new job in this area. It pays ( sometimes literally ) to know how to make your choices. So here are my thoughts.

MATLAB for Computer Vision

Until recently Computer Vision was a research area in its infancy. People who worked in Computer Vision were mostly in the academia or research labs. Their tool of choice was MATLAB, and for the longest time OpenCV paled in comparison to what MATLAB and its community had to offer. In the last 7 years, the tides have turned. Here are some of the pros of using MATLAB.

Why Should You Use MATLAB for Computer Vision : The Pros

1. Powerful matrix library : Not only do we treat an image as a multi-dimensional matrix in computer vision, we use a heavy dose of linear algebra in various algorithms. MATLAB’s linear algebra routines are very powerful and blazingly fast ( when used correctly ). Once we needed to solve a large sparse linear system as part of an algorithm. It was a one line code in MATLAB — something like X = A \ b . We were assuming that our C++ implementation will be at least 3x faster than the MATLAB implementation. The first version of our C++ implementation turned out to be slower than the MATLAB version! It took us days to match the performance of the MATLAB routine. So in MATLAB a single operator like backslash ( \ ) is sometimes a collection of powerful algorithms that MATLAB appropriately chooses for you.
2. Toolboxes : What ever your needs, there is a toolbox for that. They have an image processing toolbox, a computer vision toolbox, and a statistical and machine learning one that provide implementations of a wide variety of very useful algorithms. The functions usually provide a clean and obvious interface. Many computer vision problems are often set up as optimization problems. You are trying to maximize or minimize some objective functions under certain constraints. The optimization toolbox in MATLAB provides excellent implementations of many optimization algorithms.
3. Visualization and debugging tools : One of the joys of using MATLAB is that writing code, visualizing results, and debugging happens in one integrated environment. The environment makes you extremely productive.
4. Works with OpenCV : You can interface with OpenCV using MATLAB’s OpenCV Interface.

Why should you use OpenCV (C++) for Computer Vision : The Pros

1. It’s free! : Large parts of OpenCV are free — i.e. free as in beer, and free as in speech! You can use OpenCV freely in your commercial application, and you can view the source and fix issues if needed. You do not have to open source your project if you use OpenCV.
2. Huge optimized library : The collection of algorithms available in OpenCV dwarfs everything out there. The library is also optimized for performance. With OpenCV 3, you can use the Transparent API to easily use OpenCL compliant devices (e.g. GPU) on your machine. Many algorithms have a CUDA implementation.
3. Platforms and devices : You can obviously use OpenCV in your desktop app or as the backend of your web application. Because of its focus on performance OpenCV (C/C++) is the vision library of choice in many embedded vision applications and mobile apps.
4. Big community : There is a big community of developers (47,000 or so) that use and support OpenCV. It has been download more than 9 Million times. Unlike the MATLAB community that consists of researchers, the OpenCV community is a mix of people from many fields and industries. The OpenCV development is funded by companies like Intel, AMD & Google. Needless to say, this blog is part of this vibrant community of people who are trying to help each other.

Why should you not use OpenCV (C++) for Computer Vision : The Cons

1. Difficult for beginners : If you do not have programming experience in C++, using the OpenCV (C++) will be daunting. You are probably better off trying to use the Python instead.
2. Weak documentation : OpenCV documentation is bad. Sometimes you need to have a good understanding of the algorithm, and actually read the paper, because the documentation does not always explain what the parameters mean and how they effect the outcome. The documentation does not always come with sample code, and that makes it harder to understand. The sample code that comes with OpenCV, though very useful, is also not very well documented either. Let’s just say this is work in progress, and people are chipping in to make it better.
3. Small machine learning library : A computer vision engineer frequently needs many machine learning routines. OpenCV has a small set of machine learning algorithms compared to the choices available when you are using OpenCV (Python).
4. Visualization and debugging : Debugging and visualizing is hard in any C++ environment. This is especially true if you are coming up with a new algorithm from scratch. Occasionally I dump data to disk, and analyze it using MATLAB.

OpenCV (Python) for Computer Vision

I believe python bindings for OpenCV have contributed quite a bit to its popularity. It is an excellent choice for learning Computer Vision, and is good enough for a wide variety of real world applications. Let’s look at the pros and cons.

Why should you use OpenCV (Python) for Computer Vision : The Pros

1. Ease of use : If you are a python programmer, using OpenCV (Python) would be very easy. Python is an easy language to learn ( especially compared to C++ ). It is also an excellent first language to learn.
2. Python has become the language of scientific computing : A few years back MATLAB was called the language of scientific computing. But now, with OpenCV, numpy, scipy, scikit-learn, and matplotlib Python provides a powerful environment for learning and experimenting with Computer Vision and Machine Learning.
3. Visualization and debugging : When using OpenCV (Python) you have access to a huge number of libraries written for Python. Visualization using matplotlib is about as good as MATLAB. I find debugging code in Python easier than in C++, but it does not quite match the super-easiness of MATLAB.
4. Building web backend : Python is also a popular language for building websites. Frameworks like Django, Web2py, and Flask allow you to quickly put together web apps. It is very easy to use OpenCV (Python) along with these web frameworks. E.g. read this tutorial that explains how to turn your OpenCV code into a web api in under 10 minutes.

Why should you not use OpenCV (Python) for Computer Vision : The Cons

1. 1. Weak Documentation : As I mentioned before, OpenCV(C++) documentation is not very good. But the python documentation is even worse. A novice user is left guessing how to use certain functions. E.g. at the time of writing this post it is almost impossible to find the python documentation for cv2.Subdiv2D. You will also find fewer tutorials for OpenCV (Python).
   2. Lack of support : Companies that support OpenCV ( Intel, AMD, NVidia etc. ) have a dog in the fight when it comes to the C++ version of OpenCV. They want you to use OpenCV and to buy their hardware (CPUs/GPUs etc.) to run these algorithms. But OpenCV (Python) seems to be the proverbial red-headed step child that does not seem to get their attention. E.g. I have not yet figured out if the Transpaent API is supported in OpenCV ( Python ).
   3. Slower run time : Compared to C++, your programs in Python will typically run slower. To add an extra punch you can use the GPU ( using CUDA or OpenCL ) in OpenCV (C++) and have code that runs 10x faster than the Python implementation.
   4. OpenCV is written in C/C++ : One of the great benefits of an open source library is your ability to modify them to suit your needs. If you want to modify OpenCV, you have to modify the C/C++ source.