Dlib, when I used it for work, was a real pleasure to use. Some of its algorithms are actually state-of-the-art compared to other (often better known) open-source implementations. I would even say some of the advantages my team had over our competitors were partly due to our deeper understanding of the library ecosystem and using libraries like dlib instead of more popular default choices.

edit: grammar

Here's an interview I came across with the creator (Davis King) of dlib:

https://www.pyimagesearch.com/2017/03/13/an-interview-with-d...

Goes into some interesting details, and is a library that I'm definitely going to check out.

Does dlib use boost or lapack/scalapack underneath the hood for parallel linear algebra operations?

Have a look at this paper, not only does it support BLAS, but is also capable of rearranging matrix operations to make better use of BLAS routines.

EDIT: the paper https://jmlr.csail.mit.edu/papers/volume10/king09a/king09a.p...

Thank you. I will read the paper. I'm looking for an alternative to pytorch which works great for toy problems but starts to suffer at scale.

Dlib can use whichever BLAS library you want it to.

Second this. Used this in production to train custom detectors with headgear on.

I've added DLib usage examples in an ffmpeg video player I have on Github. Examples so far include face detection, face feature recovery, and face image collection. Super simple to use. If interested https://github.com/bsenftner/ffvideo

It is my favorite all-purpose library.

Definitely. Building it is relatively straightforward, there are many great utilities in there (not just ML stuff), the maintainer cares a great deal, and the documentation is truly exceptional (like, overall. Not just the website docs).

for real time face detection opencv actually is much faster than dlib.

https://towardsdatascience.com/face-detection-models-which-t...

Has anyone seen a way to call this from Rust?

Since it compiles a shared library, I'm assuming it's the same process as linking any other shared library in Rust

I'll take a good look at this one; as I'm currently using Python in ML, I might as well try it on C++.

How does this compare to other toolkits?

Although there is a Python frontend, Dlib is really a C++-first toolkit that can be easily integrated in to other C++ applications.

The examples are thoroughly documented, and the code itself has been a great pedagogical tool for me and others I work with. Basically, it's simple and well written.

Last time I used it, the Python package isn't shipped as a wheel, so you have to compile dlib from source using CMake on install. Makes it harder to use for Python devs.

That is correct. When one runs `pip install dlib`, only a .tar.gz file is retrieved. Then the building process takes a long while!

> Collecting dlib

> Downloading dlib-19.22.0.tar.gz (7.4 MB)

> || 7.4 MB 14.7 MB/s

> Building wheels for collected packages: dlib

I had not noticed before because I have only used dlib on Google Colaboratory, where it is thankfully pre-installed.

Nice to know there's a python wrapper, might make it easier to mock up / do initial validation before heading into the long grass.

Does anyone know Dlib workarounds/upgrades for face recognition with masks on?

Is there a good port of this for android?

Yes, an unofficial one: https://github.com/tzutalin/dlib-android