Commit Graph

333 Commits

Author SHA1 Message Date
Tadas Baltrusaitis
f8e51bca28 Simplification and speedup of landmark validation. 2017-09-01 14:58:17 +01:00
Tadas Baltrusaitis
8e4ae1ace1 Another speedup with a fortran call. 2017-09-01 11:36:26 +01:00
Tadas Baltrusaitis
e0b8e775e7 Face detection speedup through faster convolution. 2017-09-01 11:19:45 +01:00
Tadas Baltrusaitis
3fcb3ce22b Head pose estimation with multiple hypothesis re-initialization (more accurate and can start from profile). 2017-08-31 20:13:21 +01:00
Tadas Baltrusaitis
fb2ad3af7b More cleanup of head pose experiments. 2017-08-31 18:48:58 +01:00
Tadas Baltrusaitis
a4fdaa0099 Head pose results using CE-CLM. 2017-08-31 12:51:28 +01:00
Tadas Baltrusaitis
c0d6306002 In case of MTCNN pick the biggest rather than most confident face. 2017-08-31 09:45:55 +01:00
Tadas Baltrusaitis
65172b8305 Some more speedup. 2017-08-30 20:01:25 +01:00
Tadas Baltrusaitis
53a1881ede Some cleanup. 2017-08-30 15:35:03 +01:00
Tadas Baltrusaitis
0ecf4e152c A bug fix 2017-08-30 15:25:26 +01:00
Tadas Baltrusaitis
b71b2f1cfc Another performance improvement by rearranging GEMM order. 2017-08-30 15:20:56 +01:00
Tadas Baltrusaitis
e1bf2c5a39 Another slight performance increase. 2017-08-29 21:43:27 +01:00
Tadas Baltrusaitis
cc3264e749 Another small speedup. 2017-08-29 11:39:18 +01:00
Tadas Baltrusaitis
4d4f0e47e1 Another speedup. 2017-08-29 08:08:20 +01:00
Tadas Baltrusaitis
d0d8a2d762 Another speedup, by optimizing the bias addition. 2017-08-29 07:55:39 +01:00
Tadas Baltrusaitis
8e2239ad81 Speeding up sigmoid. 2017-08-28 17:05:49 +01:00
Tadas Baltrusaitis
98e9d3e6f2 Faster way to call gemm through fortran. 2017-08-28 12:11:38 +01:00
Tadas Baltrusaitis
7e4109070b Big speedup of CNN using OpenBLAS, still lots of potential for speedups 2017-08-28 10:26:24 +01:00
Tadas Baltrusaitis
2b7620e56a Adding missing OpenBLAS binaries. 2017-08-28 08:32:26 +01:00
Tadas Baltrusaitis
c0b175c73a Adding missing OpenBLAS files 2017-08-28 08:27:27 +01:00
Tadas Baltrusaitis
6414efef72 - A better way to link to OpenBLAS (using pre-compiled binaries), this leads to faster computation on Windows
- Demo script fixes
2017-08-27 15:51:40 +01:00
Tadas Baltrusaitis
f62516789c Simplifying conversion to grayscale 2017-08-26 17:01:54 +01:00
Tadas Baltrusaitis
2ea630c834 Removing debug visualization. 2017-08-26 15:51:15 +01:00
Tadas Baltrusaitis
ba6a497fcc Better output visualization in images when there are multiple faces 2017-08-26 15:47:10 +01:00
Tadas Baltrusaitis
e8e93c3f96 Incorporating MTCNN face detection with CE-CLM.
Merge remote-tracking branch 'origin/feature/mtcnn_face_det' into feature/CE-CLM

# Conflicts:
#	exe/FaceLandmarkImg/FaceLandmarkImg.cpp
#	exe/FaceLandmarkVid/FaceLandmarkVid.cpp
#	exe/FaceLandmarkVidMulti/FaceLandmarkVidMulti.cpp
#	exe/FeatureExtraction/FeatureExtraction.cpp
#	lib/local/LandmarkDetector/LandmarkDetector.vcxproj
#	lib/local/LandmarkDetector/LandmarkDetector.vcxproj.filters
#	lib/local/LandmarkDetector/src/LandmarkDetectorFunc.cpp
#	lib/local/LandmarkDetector/src/LandmarkDetectorParameters.cpp
#	lib/local/LandmarkDetector/src/LandmarkDetectorUtils.cpp
2017-08-26 13:23:16 +01:00
Tadas Baltrusaitis
92dde3b80f Speeding up the fully connected layer through better data layout 2017-08-24 17:00:32 +01:00
Tadas Baltrusaitis
a25da69550 A big PReLU speedup by making it in place and iterating smarter over it. 2017-08-23 22:25:28 +01:00
Tadas Baltrusaitis
a2a3bd08c2 Another optimization of moving the bias term addition to the matrix multiplication. 2017-08-23 20:43:19 +01:00
Tadas Baltrusaitis
b57bdd2127 Speeding up MTCNN a bit, through optimization of im2col and transposing the convolution multiplication so it is easier to split 2017-08-23 20:00:18 +01:00
Tadas Baltrusaitis
452d1b8008 Integrating MTCNN in all of the demos. 2017-08-18 20:57:19 +01:00
Tadas Baltrusaitis
a251415fea Starting to incorporate MTCNN 2017-08-18 10:03:01 +01:00
Tadas Baltrusaitis
a3be476d4d Parallelizing the MTCNN face detection 2017-08-17 14:17:20 +01:00
Tadas Baltrusaitis
441777c9ed Code cleanup and rearrangement. 2017-08-17 12:35:11 +01:00
Tadas Baltrusaitis
8d01ea5911 Different implementation of convolution, through matrix multiplication which is quite a bit faster. 2017-08-17 11:40:57 +01:00
Tadas Baltrusaitis
2bd085cb47 Exploring CNN optimization 2017-08-16 17:14:00 +01:00
Tadas Baltrusaitis
d062ab690d Final bit of simplification and fixing an issue with RNet and ONet. 2017-08-15 21:53:25 +01:00
Tadas Baltrusaitis
f4e73cc23f Simplification of flow 2017-08-15 18:06:12 +01:00
Tadas Baltrusaitis
f8519bf385 Simplifying CNN code, plus max-pooling bug fix. 2017-08-15 17:57:25 +01:00
Tadas Baltrusaitis
0c2137d950 Adding bounding box correction and an addition to non maximum supression. 2017-08-15 11:44:05 +01:00
Tadas Baltrusaitis
dd17680e2c RNet and ONet application. 2017-08-15 10:13:08 +01:00
Tadas Baltrusaitis
02a66563e1 Fixing max-pooling for different stride sizes. 2017-08-15 08:55:37 +01:00
Tadas Baltrusaitis
1d76c96cc8 Finishing the PNet step and preparing output for RNet. 2017-08-14 22:07:21 +01:00
Tadas Baltrusaitis
0f94d53d9d Helper functions for MTCNN implementation. 2017-08-14 16:23:44 +01:00
Tadas Baltrusaitis
3f2e6f228a C++ testing on 300W test set. 2017-08-11 15:01:32 -04:00
Tadas Baltrusaitis
6dbed701bc Fix with CNN inference sizes. 2017-08-11 14:56:45 -04:00
Tadas Baltrusaitis
5588742e32 300W experiment with C++ data and sparse response maps. 2017-08-11 13:49:55 -04:00
Tadas Baltrusaitis
bfb9129b31 Adding sparse response map results on 300VW. 2017-08-11 11:04:36 -04:00
Tadas Baltrusaitis
abcc91364e Towards MTCNN, implementing PNet 2017-08-10 15:30:35 -04:00
Tadas Baltrusaitis
13e42a70cf Some more work on MTCNN. 2017-08-09 16:16:31 -04:00
Tadas Baltrusaitis
7fe7bb4904 Towards MTCNN c++ implementation. 2017-08-09 15:04:14 -04:00