4dface Versions Save

Updated eos to the latest master (v0.15.0), and updated the 4dface app code accordingly.

We make a few unnecessary cv::Mat to eos::core::Image<T> conversions, but overall, it seems to run faster, thanks to all the updates in eos. (There are still many things left to "modernise", but this first update seems to hopefully run fine.)

Make sure to clone the repository with git clone --recursive, or, if you're updating, run git submodule update --init --recursive.

As always, for people doing any serious computer vision work or research, I recommend directly working with the eos library, which is in active development and frequently updated.

There's also an updated precompiled binary for Windows available. No guarantees that it'll work on many other systems though, so these are not supported. The file opencv_ffmpeg341_64.dll needs to be copied to the bin/ directory alongside the other dll's.

Windows binaries are compiled for Windows 10 64bit, Visual Studio 2017 15.6.4.

If you use the binaries, you also have to download the landmark detector model (85 MB) and put it in the share/ directory: Download face_landmarks_model_rcr_68.bin.

This release contains just minor changes from v0.1.2, but the large improvements of the last two versions definitely warrants a bump to v0.2.0!

The changes since the last release are:

• Updated the eos submodule to v0.9.1. This mainly results in a large speed-up of the shape fitting.
• Replaced the rather large 3DMM fitting code with the fit_shape_and_pose function. It's essentially doing more or less the same as before, but now encapsulated in a library function.
• 5 PCA coefficients will be fitted by default, and the 3DMM fitting will run for 3 iterations. For a very small speed-up, it runs even quite well with only 1 iteration. If it's run for more than 3 iterations, results improve somewhat, but the differences are quite small.
• Added wireframe output

Make sure to clone the repository with git clone --recursive, or, if you're updating, run git submodule update --init --recursive.

Also there's some precompiled binaries available! No guarantees that they'll work on many other systems though, so these are not supported.

Windows binaries are compiled for Windows 8.1 64bit, Visual Studio 2015. Linux binaries are compiled for Linux Mint 18, based on Ubuntu 16.04, gcc-5.4.

If you use the binaries, you also have to download the landmark detector model (85 MB) and put it in the share/ directory: Download face_landmarks_model_rcr_68.bin.

Added edge fitting from latest eos devel branch. This now fits the face model to the occluding contour as well, which results in a quite noticable fitting quality improvement!

A minor update:

• Updated eos to the latest version, v0.8.0.
• Replaced the nonlinear camera estimation algorithm with the new linear scaled orthographic projection estimation algorithm.

Initial release of the 4dface demo app. It provides good tracking and 3D fitting, but is a bit slow, due to recent additions.

Note: The CMake script will download the landmark detection model from patrikhuber.ch, which might take a bit (84.6 MB).