this started out as a try to improve opencv's lbph facereco,
but meanwhile it morphed into a testbed for comparing preprocessing/extraction/classification methods
it all builds on top of opencv30 .
the pipeline is:
(preprocessing) -> extractor -> filter -> classifier (or verifier)4 projects in here:
face3> duel -o
[extractors] :
Pixels( 0) Lbp( 1) Lbpu( 2) Lbp_P( 3) Lbpu_P( 4)
Ltp( 5) TPLbp( 6) TpLbp_P( 7) FPLbp( 8) FpLbp_P( 9)
MTS(10) MTS_P(11) BGC1(12) BGC1_P(13) COMB(14)
COMB_P(15) GaborGB(16) Sift(17) Grad(18) Grad_P(19)
GradMag(20) GradMagP(21) GradBin(22) HDGRAD(23) HDLBP(24)
HDLBP_PCA(25) PCASIFT(26) PNET(27) CDIKP(28) LATCH2(29)
DAISY(30) PATCH(31)
[filters] :
none( 0) MEAN( 1) HELL( 2) POW( 3) SQRT( 4)
WHAD_( 5) WHAD4( 6) WHAD8( 7) RP( 8) DCT_( 9)
DCT2(10) DCT4(11) DCT6(12) DCT8(13) DCT12(14)
DCT16(15) DCT24(16)
[classifiers] :
N_L2( 0) N_L2SQR( 1) N_L1( 2) N_HAM( 3) H_HELL( 4)
H_CHI( 5) COSINE( 6) KLDIV( 7) SVM_LIN( 8) SVM_POL( 9)
SVM_RBF(10) SVM_INT(11) SVM_INT2(12) SVM_HEL(13) SVM_HELSQ(14)
SVM_LOW(15) SVM_LOG(16) SVM_KMOD(17)SVM_CAUCHY(18) SVM_MULTI(19)
PCA(20) PCA_LDA(21) LDA(22) MLP(23) KNN(24)
RTREE(25)
(test/train times are averaged per fold) ------------------------------------------------------------------------------ att_faces : 10 fold, 39 classes, 390 images, retina ------------------------------------------------------------------------------ [extra] [filt] [class] [f_bytes] [hit] [miss] [acc] [t_train] [t_test] Pixels none N_L2 12100 709 32 0.957 0.000 0.641 Pixels none SVM_POL 12100 722 19 0.974 5.269 0.748 Pixels none PCA_LDA 12100 689 52 0.930 16.371 0.236 * L2 Pixels none PCA_LDA 12100 696 45 0.939 15.927 0.284 * mahalanobis Lbp none H_CHI 65536 700 41 0.945 -0.001 4.261 Lbp none H_HELL 65536 727 14 0.981 0.001 5.701 Lbp none SVM_LIN 65536 727 14 0.981 10.661 2.792 Lbpu none MLP 15360 740 1 0.999 80.554 0.250 Lbpu_P DCT6 SVM_LIN 24000 733 8 0.989 2.274 0.753 Lbpu_P DCT6 MLP 24000 740 1 0.999 144.090 0.429 LQP DCT8 MLP 32000 739 2 0.997 191.898 0.988 * LQPDisk R 5 P 17 C 220 E 3 MTS_P none SVM_INT2 11136 731 10 0.987 0.704 0.178 COMB_P HELL SVM_INT2 66816 732 9 0.988 9.312 1.009 TpLbp_P DCT8 SVM_INT2 32000 730 11 0.985 2.399 0.483 FpLbp_P none SVM_INT2 11136 731 10 0.987 0.796 0.186 FpLbp_P none MLP 11136 738 3 0.996 92.692 0.133 HDGRAD DCT24 SVM_LIN 96000 662 79 0.893 15.342 3.709 HDLBP DCT24 SVM_INT2 96000 647 94 0.873 11.912 1.336 Sift HELL SVM_INT2 492032 716 25 0.966 101.979 11.528 Sift DCT24 SVM_LIN 96000 694 47 0.937 17.222 4.562 Grad_P MEAN SVM_LIN 32016 731 10 0.987 4.156 1.505 GradMag none MLP 23552 694 47 0.937 165.413 0.549 GradMagP WHAD8 SVM_LIN 32000 733 8 0.989 3.822 1.196 PCASIFT none SVM_LIN 25600 658 83 0.888 2.689 1.041 GaborGB MEAN SVM_LIN 36864 727 14 0.981 4.494 1.490 PNET none SVM_LIN 23040 737 4 0.995 2.444 0.948 PNET none MLP 23040 741 0 1.000 113.726 0.461 CDIKP DCT8 SVM_INT2 32000 697 44 0.941 2.438 0.487 ------------------------------------------------------------------------------ yale_cropped : 10 fold, 15 classes, 165 images, retina ------------------------------------------------------------------------------ [extra] [filt] [class] [f_bytes] [hit] [miss] [acc] [t_train] [t_test] Pixels none N_L2 12100 247 53 0.823 0.000 0.102 Pixels none SVM_POL 12100 268 32 0.893 0.841 0.120 Pixels none PCA_LDA 12100 256 44 0.853 2.388 0.037 * L2 Pixels none PCA_LDA 12100 268 32 0.893 2.321 0.029 * mahalanobis Lbp none H_CHI 65536 273 27 0.910 0.001 0.647 Lbp none H_HELL 65536 293 7 0.977 0.001 1.002 Lbp none SVM_LIN 65536 296 4 0.987 1.577 0.158 Lbpu DCT2 MLP 8000 296 4 0.987 14.785 0.011 Lbpu none SVM_LIN 15360 296 4 0.987 0.235 0.028 Lbpu_P DCT8 SVM_LIN 32000 296 4 0.987 0.535 0.056 MTS_P none SVM_INT2 11136 294 6 0.980 0.111 0.011 COMB_P HELL SVM_INT2 66816 293 7 0.977 1.321 0.144 TpLbp_P DCT8 SVM_INT2 32000 296 4 0.987 0.377 0.073 FpLbp_P none SVM_INT2 11136 294 6 0.980 0.113 0.023 FpLbp_P none MLP 11136 296 4 0.987 12.102 0.019 HDGRAD DCT24 SVM_LIN 96000 296 4 0.987 2.247 0.212 HDLBP DCT24 SVM_INT2 96000 291 9 0.970 2.222 0.246 Sift HELL SVM_INT2 492032 289 11 0.963 14.645 1.683 Sift DCT24 SVM_LIN 96000 286 14 0.953 2.804 0.255 Grad none SVM_LIN 11776 296 4 0.987 0.201 0.029 Grad_P MEAN SVM_LIN 32016 295 5 0.983 0.674 0.080 GradMag none MLP 23552 283 17 0.943 35.060 0.104 GradMagP WHAD8 SVM_LIN 32000 296 4 0.987 0.684 0.083 PCASIFT none SVM_LIN 25600 285 15 0.950 0.422 0.067 GaborGB MEAN SVM_LIN 36864 293 7 0.977 0.987 0.104 PNET none SVM_LIN 23040 294 6 0.980 0.430 0.069 PNET none MLP 23040 297 3 0.990 18.541 0.060 CDIKP DCT8 SVM_INT2 32000 283 17 0.943 0.446 0.082 ------------------------------------------------------------------------------ lfw3d-aligned: 10 fold, 50 classes, 500 images, retina ------------------------------------------------------------------------------ [extra] [filt] [class] [f_bytes] [hit] [miss] [acc] [t_train] [t_test] Pixels none N_L2 12100 371 579 0.391 0.000 1.200 Pixels none SVM_POL 12100 578 372 0.608 9.938 1.300 Pixels none PCA_LDA 12100 647 303 0.681 31.688 0.319 * L2 Pixels none PCA_LDA 12100 667 283 0.702 30.349 0.485 * mahalanobis Lbp none H_CHI 65536 480 470 0.505 0.002 6.909 Lbp none H_HELL 65536 590 360 0.621 -0.000 9.209 Lbp none SVM_LIN 65536 693 257 0.729 19.759 6.148 Lbpu none SVM_LIN 15360 680 270 0.716 2.685 1.032 Lbpu_P DCT8 SVM_LIN 32000 672 278 0.707 6.266 2.538 Lbpu_P DCT4 MLP 16000 817 133 0.860 229.538 0.680 * LQPDisk R 5 P 17 C 220 E 3 LQP DCT8 MLP 32000 813 137 0.856 235.830 1.464 MTS_P none SVM_INT2 11136 691 259 0.727 1.175 0.321 COMB_P HELL SVM_INT2 66816 708 242 0.745 16.009 1.757 TpLbp_P DCT8 SVM_INT2 32000 705 245 0.742 4.113 0.851 FpLbp_P none SVM_INT2 11136 691 259 0.727 1.166 0.291 FpLbp_P none MLP 11136 755 195 0.795 175.443 0.301 HDGRAD DCT24 SVM_LIN 96000 810 140 0.853 31.486 9.809 HDLBP DCT24 SVM_INT2 96000 756 194 0.796 26.997 3.022 Sift HELL SVM_INT2 492032 759 191 0.799 136.982 15.049 Sift DCT24 SVM_LIN 96000 732 218 0.771 26.509 8.290 Grad_P MEAN SVM_LIN 32016 693 257 0.729 7.148 2.639 GradMag none MLP 23552 703 247 0.740 295.474 1.147 GradMagP WHAD8 SVM_LIN 32000 734 216 0.773 7.123 2.594 PCASIFT none SVM_LIN 25600 750 200 0.789 4.883 2.089 GaborGB MEAN SVM_LIN 36864 720 230 0.758 8.862 3.098 PNET none SVM_LIN 23040 756 194 0.796 4.783 1.884 PNET none MLP 23040 852 98 0.897 161.164 0.846 CDIKP DCT8 SVM_INT2 32000 658 292 0.693 4.398 0.806