Ask HN: What are the best resources to learn computer vision?

I think the question is a little too unspecific for there to be a good answer. The field is vast and depending on which thing in computer vision you want to tackle the best learning paths may vary greatly. Just to give a bit of an overview:

Before the Deep Learning Craze started in 2011 more classical Machine Learning techniques were used in CV: Support Vector Machines, Boosting, Decision Trees, etc..

These were (and still are!) used as a high level component in areas like recognition, retrieval, segmentation, object tracking.

But there's also a whole field of CV that doesn't require Machine Learning learning at all (although it can benefit from it in some cases). This is typically the area of geometrical CV, like SLAM, 3D reconstruction, Structure from Motion and (Multi-View) Stereo, anything generally where you can write a (differentiable) model of reality yourself using hand-coded formulas and heuristics and then use standard solvers to obtain the model parameters given the data.

Whenever it's too hard to do that (for example trying to recognize many different things in images) you need a data-driven / machine learning approach where the computer comes up with the model itself after seeing lots of training examples.

As for resources the other answers are already giving a great overview. Use Karpathy's course for an intro to Deep Learning for CV but don't expect it to be comprehensive in terms of giving you an overview of CV.

Learn OpenCV for more low level, non-ML and generally more "old-school" Computer Vision.

A personal recommendation of mine is http://www.computervisionblog.com/ by Tomasz Malisiewicz. It's an excellent resource if you want to get an overview of what's happening in the field.

A lot of real world computer vision is implemented on embedded devices with limited computational resources (ARMs, DSPs, etc.) so understanding how a lot of commonly used algorithms can be efficiently implemented in embedded systems is important. It is possibly a way for you to jump the gap from "embedded software developer" to "computer vision engineer". Also keep in mind that in many companies a "computer vision engineer" is fundamentally a different beast from a "software developer". A CV engineer creates software but the emphasis tends to be more on systems and is not 100% about software. This will vary a lot by company but if you're working with prototype hardware you will need to get at least a working knowledge of optics.

Fun and trendy though it may be, I would not focus on deep learning / convolutional neural networks to start off. Deep learning is a small subset of computer vision. I would focus more on understanding the basics of image processing, camera projection geometry, how to calibrate cameras, stereo vision, and machine learning in general (not just deep learning). Working with OpenCV is a good place to start for all of these topics. Set yourself a project with tangible goals and get to work.

Great points.

I would argue Kinetic or Geometrical Computer Vision problems, things like Tracking, Mapping, Reconstruction, Depth Estimation are best suited for the classical approaches like VO, SFM/MVS, SIFT/SURF, HOG etc... and are a separate category of CV problems than object recognition/detection/segmentation - much more capable of being done with ML because dimensionality is reduced.

But there's also a whole field of CV that doesn't require Machine Learning learning at all (although it can benefit from it in some cases).

In fact, Machine Learning has made almost no progress on most of what you mention, specifically SLAM and Multi-View Stereo. It takes completely rethinking how those are done when they are approached from the Deep Learning perspective.

Surprised nobody has posted http://course.fast.ai/ yet. I've been following along with it so far for the first 4 lessons and it has been extremely helpful in understanding how deep learning works from the perspective of someone who did not have much of any related baseline knowledge except how to program. Jeremy is an excellent practical teacher.

For classical (non-machine learning) material, there's the nice text by Ballard and Brown: http://homepages.inf.ed.ac.uk/rbf/BOOKS/BANDB/bandb.htm

Even if you end up using neural nets, understanding how to think about the problems is useful.