DEST v0.8

We are happy to announce v0.8 of DEST bringing the following improvements

• New tracker files with increased accuracy while reducing disk space.
• Tracker evaluation now prints more statistics including a error histogram.
• Removed obsolete training parameters.

Pre-trained Trackers

dest_tracker_VJ_ibug.bin

• Trained on afw, helen, ibug, lfpw face data sets containing over 3200 images (doubled through mirroring).
• Annotated by 68 landmarks.
• Initial rectangles provided by OpenCV Viola Jones using dest::face::FaceDetector using haarcascade_frontalface_alt2.xml

Training

• Trained on: afw, helen, ibug, lfpw
• Rectangles: OpenCV VJ classifier_frontalface_alt2.xml

dest_train --load-mirrored --load-max-size 640 --rectangles train\rectangles2.csv 
--create-num-shapes 60 --train-num-pixels 600 --train-num-splits 40 --train-max-depth 5 
--train-learn 0.15 train

Loading ibug database. Found 3283 candidate entries.
Successfully loaded 5836 entries from database.
Creating training samples.
Number shapes per image       60
Linear weight range           [0.65,0.8]
Include mean shape            true
Starting to fit tracker on 355996 samples.
Number of cascades            10
Number of trees               500
Maximum tree depth            5
Random pixel locations        600
Random split tests            40
Random pixel expansion        0.05
Exponential lambda            0.1
Learning rate                 0.15
Building cascade 1
Building cascade 2
Building cascade 3
Building cascade 4
Building cascade 5
Building cascade 6
Building cascade 7
Building cascade 8
Building cascade 9
Building cascade 10
Saving tracker to dest.bin

Benchmark

dest_evaluate.exe -t dest.bin --load-max-size 640 
-r helen\testset\rectangles2.csv helen\testset

Loading ibug database. Found 330 candidate entries.
Successfully loaded 305 entries from database.
Average normalized error:               0.0370472
Stddev normalized error:                0.0366681
Median normalized error:                0.0259357
Worst normalized error:                 0.540997
Error histogram (bin size 0.05):
0.025/78.0% |**********************************************
0.075/16.2% |*********
0.125/4.0%  |**
0.175/1.1%  |
0.225/0.4%  |
0.275/0.1%  |
0.325/0.1%  |
0.375/0.1%  |
0.425/0.0%  |
0.475/0.0%  |
0.525/0.0%  |
0.575/0.0%  |
0.625/0.0%  |
0.675/0.0%  |
0.725/0.0%  |
0.775/0.0%  |
0.825/0.0%  |
0.875/0.0%  |
0.925/0.0%  |
0.975/0.0%  |
 >1.0/0.0%  |

DEST v0.7

We are happy to announce v0.7 of DEST bringing the following improvements

• switched from GPL to BSD license.
• re-trained tracker files with increased accuracy.
• training now supports nested directories.
• importing databases now searches for different image types.

Pre-trained Trackers

dest_tracker_VJ_HELEN.bin
Trained on the ibug annotated HELEN face data set containing 2000 images (doubled through mirroring) annotated with 68 landmarks. Initial rectangles provided by OpenCV Viola Jones using dest::face::FaceDetector.

Training configuration

dest_train --load-mirrored --load-max-size 640 --rectangles helen\trainset\rectangles.csv --create-num-shapes 40 --train-num-pixels 600 --train-num-splits 40 --train-max-depth 6 helen\trainset

Loading ibug database. Found 2000 candidate entries.
Successfully loaded 4000 entries from database.
Creating training samples.
Number shapes per image                 40
Number of transforms per shape          1
Random rotate angle range               [0,0]
Random scale factor range               [0.85,1.15]
Random translate x range                [-10,-10]
Random translate y range                [-10,10]
Use linear combination                  true
Starting to fit tracker on 158400 samples.
Number of cascades            10
Number of trees               500
Maximum tree depth            6
Random pixel locations        600
Random split tests            40
Random pixel expansion        0.05
Exponential lambda            0.1
Learning rate                 0.08

Performance on HELEN test data set

dest_evaluate.exe -t dest.bin -r helen\testset\rectangles.csv helen\testset

Loading ibug database. Found 330 candidate entries.
Successfully loaded 330 entries from database.
Average normalized error: 0.0411167

DEST v0.6

Initial release of DEST featuring

• A generic framework for learning arbitrary shape transformations.
• A lightning fast landmark alignment module.
• State of the art performance and accuracy.
• Pre-trained trackers for a quick start.
• Cross platform minimal disk footprint serialization.
• Built in support for IMM and ibug annotated face database import.

Pre-trained Trackers

dest_tracker_VJ_HELEN.bin
Trained on the ibug annotated HELEN face data set containing 2000 images (doubled through mirroring) annotated with 68 landmarks. Initial rectangles provided by OpenCV Viola Jones using dest::face::FaceDetector.

Training configuration

Loading ibug database. Found 2000 candidate entries.
Successfully loaded 4000 entries from database.
Creating training samples.
Number shapes per image                 20
Number of transforms per shape          10
Random rotate angle range               [0,0]
Random scale factor range               [0.85,1.15]
Random translate x range                [-10,-10]
Random translate y range                [-10,10]
Use linear combination                  true
Starting to fit tracker on 792000 samples.
Number of cascades            10
Number of trees               500
Maximum tree depth            5
Random pixel locations        400
Random split tests            20
Random pixel expansion        0.05
Exponential lambda            0.1
Learning rate                 0.08

Performance on HELEN test data set

Loading ibug database. Found 330 candidate entries.
Successfully loaded 330 entries from database.
Average normalized error: 0.0451457