Of course my initial reaction is no, but I think there's some important discussion to be had in my naive understanding of this topic.
Weka (http://sourceforge.net/projects/weka/) and Ilastik (http://ilastik.org/) are two of several applications that nicely combine machine learning and image processing principles into a graphical application that is approachable to novices. These tools are general enough that we an no longer write them off to niche domainware. Using some of these tools, I get a strong black box vibe. The documentation and source code are open, so this is very much due to my own ignorance; however, I think many users are inclined to play with these programs until they get the results that they want without understanding the intricacies under the hood. This is of course a double-edged sword, but IMO is mostly good from a productivity standpoint.
Juxtapose this to core image processing libraries with intuitive API's like scikit image, ImageJ's API, those bundled with Matlab etc.... These libraries of course are pivotal for creating such applications, but I wonder in the next ten years what image processing will look like. Will everyone go their favorite ALL-IN-ONE super application, or is the slick and simple appeal of these fundamental libraries going to encourage more users to get their hands dirty so to speak? What do you guys think?
I'm also interested in assimilating a list of strengths and (moreso) weaknesses of the aformentioned blackbox, super tools. This is for a methods paper we're assimilating and I'd really like to provide that story in an honest way.
Thanks
I also worry about over abstraction, especially for teaching, where CV becomes a series of function calls or a tool that hides all the algorithmic details - on the other hand, that has been a goal and it does allow for quick exploration of meta-algorithms and CV concepts and theory. I think as long as the CV developer continues to work both bottom up from fundamental algorithms and transform methods and top-down with tools, APIs, and CV languages, then it's great. One trend I see, that goes beyond APIs and tools is the innovation of CV languages like Halide (http://halide-lang.org/) and the development of standards for CV co-processing like OpenVX. Both to me indicate that CV is maturing far beyond a collection of algorithms bundled in a library, but on the other hand, that's what got us to this point and I still very much use tools like OpenCV myself, or build algorithms ground up in C/C++ and even with FPGA state machines. The languages and co-processors are where I see the future myself.
I also worry about over abstraction, especially for teaching, where CV becomes a series of function calls or a tool that hides all the algorithmic details - on the other hand, that has been a goal and it does allow for quick exploration of meta-algorithms and CV concepts and theory. I think as long as the CV developer continues to work both bottom up from fundamental algorithms and transform methods and top-down with tools, APIs, and CV languages, then it's great. One trend I see, that goes beyond APIs and tools is the innovation of CV languages like Halide (http://halide-lang.org/) and the development of standards for CV co-processing like OpenVX. Both to me indicate that CV is maturing far beyond a collection of algorithms bundled in a library, but on the other hand, that's what got us to this point and I still very much use tools like OpenCV myself, or build algorithms ground up in C/C++ and even with FPGA state machines. The languages and co-processors are where I see the future myself.
I also agree that while the 'black box super tools' bring CV and ML to the 'masses' they also make for some very poor scientific research submissions (I see this when reviewing for conferences and journals) that figure out that they will just run their dataset on WEKA and if they get 85+% classification accuracy, then they have found a major breakthrough. As the person before me said, a 'basic' understanding of CV and ML/DM algorithms is needed to be able to use them properly in research (and industry) and this seems to be going away with the use of such convenient and helpful tools.
I would love to have everything in one super tools (since WEKA and OpenCV for example) greatly differ and they while they complement each other, it takes a bit to get them to talk to each other nicely. While working in interdisciplinary research, I have found that the 'super tools' seem to be getting more attention since it speeds things up rather than learning a language. And with proper education of the basics (WEKA has good pointers to the actual papers behind the algorithms they use) using 'super tools' should be a way to go for certain types of research.
Thanks guys, these answers were quite insightful. Juan, OpenCV provides a tools like WEKA? I had naively thought it was just a library.
What I meant is that it is a tool since it has many algorithms together under a single package (library). It does not have a nice GUI like WEKA to do stuff. But I still consider both tools, one for DM/ML and the other one for CV.
Weka may be a super-tool, but I think it also benefits from its incorporation into standard image processing libraries. For example take a look at http://fiji.sc/Trainable_Weka_Segmentation This plugin to ImageJ view and manipulate weka results. This helps the novice user review, and more importantly question, the results that weka gives by giving them tools that don't require learning a language.
Even if the "supertools" like weka help creating novel algorithms and methods for new applications, they have to be implemented later in a real system.
We will still need efficient API implementations for real-time embedded processing...
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing