Releases: TadasBaltrusaitis/OpenFace
Releases · TadasBaltrusaitis/OpenFace
OpenFace 2.2.0
Update of OpenFace to version 2.2.0, a number of bug fixes and library updates
Change log:
- Moving to C++17. This means that the code can only be build using C++17 compilers (e.g. g++ >8 and Visual Studio 2017, clang > 5), fixing related bugs - (#698, #629, #641)
- Removing an explicit dependency on boost (all the filesystem operations are performed using std::filesystem or boost::filesysteme). If boost is available it will used boost:filesystem, otherwise std::filesystem (this requires C++17).
- Visual Studio 2017 is now the main version for Visual Studio builds, VS 2015 is no longer supported
- Updating OpenCV to 4.1 version (#511)
- Fixing a bug with output images when using GUI (#694)
- Updating RAWImage - #609, so it can be initialized from System.Drawing.Bitmap directly
- Fixing overlap detection for multi face tracking: (#693)
OpenFace 2.1.0
- Removing TBB dependency. TBB can still be used, but is called through OpenCV (
#241) - Updating OpenBLAS version in 64-bit Windows builds (should lead to significant performance gains on that architecture)
- Allow specifying image output format (for both aligned images and visualized images). See wiki for more details
- Removing unused -q flag (#593)
OpenFace 2.0.6
OpenFace 2.0.5
Changelog:
- Fixing a crash in OpenFaceOffline in processing images without any faces present (#488)
- Fixing a crash in OpenFaceDemo that was caused by memory management issues in wrapper between C# and C++ (#453)
- Fixing waitKey being called even when no output is visualized
- Removing AVX requirements on Windows code (will make it more compatible with some older machines)
- Fixing bbox directory issue (#539)
OpenFace 2.0.4
Changelog:
- Fixing a bug that sometimes caused a crash when the face left the screen
- Updating readme for OSX to do with memory issues in OpenBLAS
- Hard stop in CMake if Boost or OpenCV not found
- Fixing Haar detector copy in CMake (mentioned in #509)
- Getting ready for OpenCV 4.0 (thanks #511)
- Updating wiki on Mac installation and possible OpenBLAS issues
OpenFace 2.0.3
A lot of house-keeping: updating libraries and extra testing on various Ubuntu versions.
Changelog:
- Big re-org and cleanup of CMake (thanks aleh-arol - https://github.com/aleh-arol and severin-lemaignan - https://github.com/severin-lemaignan)
- CMake is now only for Ubuntu and OSX, while for windows builds use Visual Studio 2015
- Updating dlib to 19.13
- Bringing up to recent build that has fewer clashes with OpenBLAS and boost
- Now the code will require a separate build to link to on OS X and Ubuntu (this will allow users to re-use their own dlib builds and will increase stability)
- Instructions and testing for Ubuntu 14, 16, and 18
- Fixing boost and dlib clashes, that resulted in requiring a certain version of boost, now most versions of boost should work
- OpenBLAS include fixes with testing on Ubuntu 14.04, 16.04, 18.04, and OS X
- Some of the issues were caused by Ubuntu 14.04 openblas install not containing headers
- Some CNN util cleanup
- Cleaning up includes in Cpp interop code
- Bug fixes:
- Really short executables using FeatureExtraction
- Option to process .webm files in the GUI
OpenFace 2.0.2
Mainly stability fixes:
- Cmake fixes and updates for OS X at Travis CI
- Threading issues resolved on certain versions of TBB on OS X and Ubuntu
- Fixing numerical stability bug with rotation matrices
- Fixing issue of OpenBLAS linking for OS X (default is to link to Accelerate)
- Addition of a docker image (thanks Edgar Aroutiounian)
OpenFace 2.0.1
Bug fixes:
- Fixing output video and image writing in Windows GUI
- Ability to us multi face tracking with x86 architectures (not running out of memory)
- Smaller memory footprint when using multiple CE-CLM models
Code cleanup:
- CEN response map computation path simplification
Features:
- Ability to tell OpenFace not to output aligned faces when detection unreliable
OpenFace 2.0.0
A lot of changes:
- A new facial landmark detection model - CE-CLM. Much more accurate model that is able to cope much better with profile faces
- Training code for new models
- Slightly slower than previous model, but previous models can still be used as well
- A new CNN based face detector (MTCNN) that can detect faces in much more difficult scenarios and even at profile faces!
- Both combined lead to better results on landmark detection, head pose estimation, gaze, and AUs
- GUI better support for multi-face images (correct visualization, output fixes with face id)
- GUI selection of face detector and landmark detector
- bug fix with memory allocation in visual C++ bindings
- Experimental evaluation on 300VW data
- Experimental evaluation on IJB-FL data
- Matlab hierarchical fitting has much cleaner and clearer code
- A lot of code optimizations
- Threading based input and output
OpenFace 1.0.0
Changelog:
- AU visualization module
- Stability fixes when AU or gaze models unavailable due to licensing
- Multi-PIE trained models added
- Bug fix with visualizing when faces are not detected in first frame