Hacker News Reader

Comment by a_d | Hacker News Reader

Is OpenCV (traditional CV technique) better to use or Deep Learning based approach? Has anyone done a comparison of the two approaches? The obvious flaw with deep learning is that it requires large labeled data sets - but assuming that is available, which one is more accurate at object detection (hotdog or not), detecting features on an image (faces, manufacturing defects)?

View on HN

For the widely used handwriting dataset, MNIST, there is a site that actively tracks the best algorithms (although I'm not sure when it was last updated): http://rodrigob.github.io/are_we_there_yet/build/classificat...

You'll notice that all of the top contenders use Neural Networks, but I would bet that many of them use at least some traditional CV techniques to transform the images at various steps. That said, many of the more modern deep learning approaches are ditching CV altogether, just feeding in raw pixels without any normalization or transformation, leaving fewer parameters to tweak.

I use OpenCV for reading and writing real-time video streams (like webcams or video frames) for my hobby computer vision projects, but tend to use other ML or image-specific libraries for actual processing. The cascade classifiers in OpenCV are okay, but it isn't too difficult to set up something comparable in scikit-learn that is more modern and robust. Though a bit less performant if you do need real-time response though.

Best example that I have is a pulse rate detector that I put together, that uses OpenCV for video frame extraction & display but bare numpy/scipy for the rest.

https://github.com/thearn/webcam-pulse-detector

Reply on news.ycombinator.com