This project provides SVM training, detection, and (adaptive) tracking via libraries and small demo applications.
- Create new project directory:
$ mkdir AdapTrack - Enter the new directory:
$ cd AdapTrack - Clone the repository (e.g.
$ git clone https://github.com/ex-ratt/AdapTrack.gitor using zip download), results in directory namedAdapTrack - Create build directory next to
AdapTrack:$ mkdir build - Change to build directory:
$ cd build - Build project:
$ cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=../install ../AdapTrack/ - Compile the project:
$ make - Install the libraries and binaries:
$ make install
- Create new project directory:
$ mkdir AdapTrack - Enter the new directory:
$ cd AdapTrack - Clone the repository (e.g.
$ git clone https://github.com/ex-ratt/AdapTrack.gitor using zip download), results in directory namedAdapTrack - Create build directory next to
AdapTrack:$ mkdir build - Change to build directory:
$ cd build - Build Eclipse project:
$ cmake -G"Eclipse CDT4 - Unix Makefiles" -D CMAKE_ECLIPSE_GENERATE_SOURCE_PROJECT=TRUE -D CMAKE_BUILD_TYPE=Release ../AdapTrack/
- libImageIO: Loading and storing of images and annotations
- libImageProcessing: Image filtering, pyramid construction, feature extraction
- libClassification: Interfaces for binary and probabilistic classifiers and trainers, support vector machine, kernels
- libSvm: Wrapper around libSVM that implements the classifier trainer interfaces of libClassification
- libDetection: Sliding-window-based detector and trainer
- libTracking: Adaptive model-free short-term single-target tracker and adaptive particle-filter-based multi-target tracker
Some of the applications need annotations and configuration files. You can find examples of these files in the resources directory and in the description below.
Trains and tests detectors that are based on the sliding window technique, aggregated features, and a linear support vector machines.
./DetectorTrainer train DIRECTORY IMAGES SETCOUNT [FEATURECONFIG TRAININGCONFIG]
- DIRECTORY: directory that should be created (or re-used) and stores the config files and trained SVMs
- IMAGES: path to an XML file with image names and annotations, created with dlib's imglab tool
- SETCOUNT: number of subsets to use for cross-validation (if set to 1, all images are used to create a single detector, so there won't be cross-validation)
- FEATURECONFIG: configuration file containing the feature parameters, see below for an example (only necessary if detector directory does not exist)
- TRAININGCONFIG: configuration file containing the training parameters, see below for an example (only necessary if detector directory does not exist)
Example: $ ./DetectorTrainer train detector-fhog9-4x10 annotations.xml 4 features-fhog9-4x10 training-c10
./DetectorTrainer test DIRECTORY IMAGES SETCOUNT DETECTIONCONFIG
- DIRECTORY: directory that contains config files and trained SVMs (created when training the detector)
- IMAGES: path to an XML file with image names and annotations, created with dlib's imglab tool
- SETCOUNT: number of subsets to use for cross-validation (if set to 1, all images are used to test a single detector, so there won't be cross-validation)
- DETECTIONCONFIG: configuration file containing the detection parameters, see below for an example
Example: $ ./DetectorTrainer test detector-fhog9-4x10 annotations.xml 4 detection-40x40-5
./DetectorTrainer show DIRECTORY IMAGES SETCOUNT DETECTIONCONFIG [THRESHOLD]
- DIRECTORY: directory that contains config files and trained SVMs (created when training the detector)
- IMAGES: path to an XML file with image names and annotations, created with dlib's imglab tool
- SETCOUNT: number of subsets to use for cross-validation (if set to 1, all images are used to test a single detector, so there won't be cross-validation)
- DETECTIONCONFIG: configuration file containing the detection parameters, see below for an example
- THRESHOLD: SVM threshold (optional, defaults to 0.0)
Example: $ ./DetectorTrainer show detector-fhog9-4x10 annotations.xml 4 detection-40x40-5 0.5
The app can be controlled by the following keys:
- q: quit
- other keys: progress to the next image
Feature configuration
type fhog9 ; feature type
; fhog# (Felzenszwalb's HOG variation, # = number of unsigned orientation bins)
; fpdw (features of the Fasted Pedestrian Detector in the West)
windowWidthInCells 10 ; width of the detection window in cells
windowHeightInCells 10 ; height of the detection window in cells
cellSizeInPixels 4 ; width and height of a cell in pixels
octaveLayerCount 10 ; number of image pyramid layers per octave - only used for training
Training configuration
mirrorTrainingData true ; flag that indicates whether the training data is symmetric and should be mirrored to double the amount of data
maxNegatives 30000 ; maximum number of negatives to use - 0 for unlimited
randomNegativesPerImage 20 ; initial number of negative training examples per image
maxHardNegativesPerImage 100 ; maximum number of additional hard negatives sampled per image in each bootstrapping round
bootstrappingRounds 3 ; number of bootstrapping rounds
negativeScoreThreshold -1.0 ; hard negative training examples have an SVM score of at least this value
overlapThreshold 0.3 ; maximum allowed overlap between negative training examples and annotations
C 10 ; SVM penalty multiplier
compensateImbalance true ; flag that indicates whether to compensate for data imbalance
probabilistic false ; flag that indicates whether to train a probabilistic SVM (computes and stores logistic parameters, does not influence detection)
Detection configuration
minWindowWidthInPixels 40 ; minimum width of objects to be detected
minWindowHeightInPixels 40 ; minimum height of objects to be detected
octaveLayerCount 5 ; number of image pyramid layers per octave
approximatePyramid false ; flag that indicates whether to approximate all but one layer per octave
nmsOverlapThreshold 0.3 ; maximum allowed overlap between two different detections after non-maximum suppression
Tracks a single target without prior knowledge after initialization by the ground truth.
./SingleTracker ANNOTATIONS BINS CELLSIZE TARGETSIZE PADDING ADAPTATION
- ANNOTATIONS: path to an XML file with image names and annotations, created with dlib's imglab tool
- BINS: number of unsigned orientation bins of the FHOG features
- CELLSIZE: width and height of the FHOG cells in pixels
- TARGETSIZE: size of the target in FHOG cells (larger one of width or height)
- PADDING: number of cells around the previous target position that is searched for the new position
- ADAPTATION: weight of the new SVM parameters between zero (no adpatation) and one (no memory)
Example: $ ./SingleTracker annotations.xml 9 4 10 7 0.1
The app can be controlled by the following keys:
- q: quit
- p: pause/unpause
- r: re-initialize from ground truth at next image
- other keys: progress to the next image if paused
Tracks multipe targets using a particle filter for each.
./MultiTracker VIDEO SVM CELLSIZE DETECTIONTHRESHOLD VISIBILITYTHRESHOLD
- VIDEO: camera device ID, video file, dlib annotation XML-file, or image directory
- SVM: text file that contains the SVM data (created by DetectorTrainer)
- CELLSIZE: width and height of the FHOG cells in pixels
- DETECTIONTHRESHOLD: SVM score threshold for detections to be reported
- VISIBILITYTHRESHOLD: SVM score threshold for tracks to be regarded visible
Example: $ ./MultiTracker video.avi svm-fhog9-4x10 4 1.0 -0.25
The app can be controlled by the following keys:
- q: exit
- p: pause/unpause
- r: reset the tracker
- d: show debug-output (particles and unconfirmed tracks)
- other keys: progress to the next image if paused
There are some additional useful resource files in the directory resources:
- thesis.pdf: Thesis that describes the detection and tracking approaches in chapters 4 to 6.
- svm: Few SVMs trained by the DetectorTrainer on frontal heads
- annotation: Annotations of heads that can be used to train and test head detectors
- features: Exemplary feature parameter files
- training: Exemplary training parameter files
- detection: Exemplary detection parameter files