patrikhuber / eos

A lightweight 3D Morphable Face Model library in modern C++
Apache License 2.0
1.92k stars 598 forks source link

[Suggestion] Better documentation about landmarks #246

Open ghost opened 5 years ago

ghost commented 5 years ago

Hi. I asked a question a while ago about how to generate the landmarks file. Although it was properly answered, probably my ignorance about computer vision specifics prevented me from making progress. This issue is just a suggestion: maybe, in the interest of making the system more accessible to the layman, the session "Sample code" could have a better description about what is a landmark-to-vertex file, how to generate one and what kind of format is expected by eos. This is coming from the supposition that the user has just a single photo of a face he wants to reconstruct. Eos is kind of singular in comparison to other "morphable model fitting" algoritms in that they usually make the reconstruction without the need of facial landmarks (maybe they detect them automatically). I know how to make landmark detection with opencv (according to SOME standard), but I have really no clue about how to translate them to vertices on a morphable model such as 4DFM. If the subject is too much of a common knowledge, maybe some references in the README can be very clarifying to newbies.

ghost commented 5 years ago

To complement, the replies to the following issues #69 , #31 and the guide on this question

https://github.com/patrikhuber/eos/issues/15#issuecomment-161558994

are very helpful. The major problem I still have is, after annotating the images, how can I know to which vertices those facial landmarks should correspond in a model like SFM or 4DFM. I can only think about using a 3D software like Blender to manually do the job, but this seems very error prone. In one of the answers, it is mentioned that 4DFM has landmark detection, but the whole workflow is still very unclear to me.

I think, in the end of the day, that if the section "Sample code" did not make the assumption that the user already has a landmark-to-vertex file and facial annotations, and tried to roughly explain the workflow to make the 3D reconstruction of a raw selfie in the wild, it would be much more useful.

patrikhuber commented 5 years ago

Hi David,

Thanks a lot for these suggestions. It's always good to hear about problems that users are encountering.

I am currently quite busy but will certainly take in the feedback and use it when I can make some time to work on it.

Thank you very much, and also I think many users will already just find your write-up here very useful!

ghost commented 5 years ago

Update: It turns out that dlib's facial landmark detector already uses the 68 point ibug landmark format. So there is no need to create a new "ibug_to_sfm" landmark-to-vertex file: one can just create the landmark file automaticaly using dlib and then call eos's routines. The following post explains how:

https://www.pyimagesearch.com/2017/04/03/facial-landmarks-dlib-opencv-python/

Anyway, it would be interesting to have an "ibug_to_4dface" toml file. Any chance I am missing it?

patrikhuber commented 5 years ago

dlib's facial landmark detector already uses the 68 point ibug landmark format

In my opinion that's pretty self-explanatory. It should also be documented on dlib's pages or in their documentation. If it isn't, perhaps file an issue there? But it detects 68 points, so it doesn't take much to find out that it's the 68 points from the ibug scheme. And then it should be clear that one can use the ibug_to_sfm mappings file since the landmarks are the same.

it would be interesting to have an "ibug_to_4dface" toml file. Any chance I am missing it?

What do you mean exactly with "to 4dface"? If you mean the 4DFM, then may I ask if and how you obtained it? As far as I can see you are not a licensed user but perhaps you're working for a company that we licensed it to? If you don't want to tell publicly, please drop me an email at <4dface> .

The 4DFM contains landmark definitions, they're directly included in the model file: https://github.com/patrikhuber/eos/blob/f8b235bf8da1eef947a62dc09e23ecde2c9c0f3b/include/eos/morphablemodel/MorphableModel.hpp#L387-L400

And then there's a file defining the mappings for ibug, ibug68_landmark_mappings.txt, which we make available to people on request. We may end up putting it in the eos repo at some point.

ghost commented 5 years ago

Hi Patrick,

When one learns about dlib's facial landmark detection, then the information that the format used by them is ibug's 68 points is right there. However, it took me a while to find this out, because I didn't know where to look before (I just found out about this dlib routine yesterday). Coming from python, I didn't even know what dlib was when I stumbled upon eos the first time a few months ago. So, what I've meant is that including in the Readme a hint such as "You can generate the pts landmarks file by using dlib's facial landmark detection, for example" might save some trouble for computer vision newbies, who might not be familiar with dlib.

About the 4DFM model, never mind. I saw your "4dface" repo in a glimpse and thought the model was publicly available, but now I see they are completely different things.

Anyway, I was finally able to run eos yesterday and was very impressed with the results, and now I'm studying the code more deeply. Congrats for the impressive work!

patrikhuber commented 5 years ago

"You can generate the pts landmarks file by using dlib's facial landmark detection, for example"

I see! I think that's indeed a very good suggestion, thank you very much. Something like exactly that would indeed be really good to add to the readme.

I saw your "4dface" repo in a glimpse and thought the model was publicly available, but now I see they are completely different things.

Ah, ! see :-) Yes I agree this can potentially be a bit confusing - 4dface and the patrikhuber/4dface GitHub repo has been previously a research-demo combining eos with our regression-based landmark detection - and now it's our company name! I've added a note to the top of the patrikhuber/4dface repo's readme a couple of month ago:

Note: This is a research demo from 2015/2016. 4dface has since been commercialised and an improved, up-to-date version is available on www.4dface.io. This repository also contains an older version of the morphable face model fitting library eos. eos is still in active development, head over to github.com/patrikhuber/eos for the latest version.

Maybe we should make the note bigger... :-)

I'm glad you all got it to work! And thank you very much for the congratulations!