Welcome to the cedar plugin collection

This is a collection of plugins for cedar.

Building the plugin

TODO describe how to build the plugin

Using shared buildsets

To use one of the shared buildsets follow these steps:

1. Enter your build directory. Remove your CMakeCache.txt if it exists.
2. In the build directory, type cmake -D BUILDSET=shared_buildsets/TheBuildsetYouWantToUse.cmake ..
3. Type ```make all``

Note that the name of the library generated by differnt buildsets differs. Make sure to use the appropriate shared object!

Plugin configurations

TODO find a better name for plugin configurations

Each pulgin configuration defines two things: 1. the name of the plugin being built; and, 2. what classes to build into the plugin. To declare a plugin configuration, create a buildset (put it into the shared_buildsets folder if you want others to have access to it) or edit the buildset.cmake in the root directory.

First, declare the name of the plugin configuration using

DECLARE_PLUGIN(MyPlugin)

This tells the buildsystem that the shared object (or dll on Windows, dylib on Mac OS) should be named libMyPlugin.so (Windows: MyPlugin.dll, Mac OS: libMyPlugin.dylib).

Next, you should add classes to this plugin. This is done with the ADD_TO_PLUGIN command, which offers various options. To add a specific step class, call it like this:

ADD_TO_PLUGIN(STEP some::Step)

You can also add multiple steps:

ADD_TO_PLUGIN(STEPS some::Step some::other::Step)

If you want to add all steps from a category, you can write

ADD_TO_PLUGIN(CATEGORY YourCategory)

As before, you can add multiple categories like this:

ADD_TO_PLUGIN(CATEGORIES YourCategory YourOtherCategory)

Note that you can combine several of these commands in a single ADD_TO_PLUGIN call, for example:

ADD_TO_PLUGIN(STEPS some::Step some::other::Step CATEGORY YourCategory)

Finally, you can add all classes of a kind using these commands:

ADD_TO_PLUGIN(ALL_STEPS ALL_KERNELS)

Steps included in the plugin

step	description
Algebra
CrossCorrelation	Calculates the cross correlation in different configurations.
ElementwiseDivide	To input matrices Nom and Denom will be divided elementwisely. The output is Nom/(Denom + scalar). "scalar" is a scalar and configurable parameter. This value is added to all elements in the Denom matrix to prevent dividing by zero or very small values.
PointWiseNormalization	Normalizes the entries along one dimension of a matrix.
Arrays
AttentionSlice	A step that determines the location of the maximum in an activation matrix and cuts out a region around this location from an input image.
HyperAcuteRescaling	Rescales an input matrix so that as little information as possible is lost. Note that this step's output is not properly normalized.
MaxPooling	Subsamples a matrix by taking the maximum over rectangular regions.
PadAndTranslate	A step that converts a 1D representation into a larger frame and translates the center of the 1D representation to the position given by the skalar-translation input. Borders are cyclic.
WeightedSum	A looped step that multiplies each input value with a given weight and sums them up across each input dimension. The given weight function encodes the distance from the center of the input matrix. This step is used to calculate a velocity signal dependent on the distance between a peak position and the center of the input matrix.
WeightedSumOfSlices	Calculates a weighted sum of the entries in a 3d or 4d matrix along the first dimension.
DFT
FourDimHebbMap	A step that learns an associated 2D input for each activation in a 2D field in presence of a reward signal
Integrator	A looped step that integrates nx1 inputs across time.
RewardHebbTrace	A step that learns an associated input in presence of a reward signal
Dynamics
InhibitoryNeuron	A single neuron with the dynamics tau * dv/dt = -sigmoid(s) * tanh(beta * (v - s)) - (1 - sigmoid(s)) * v, where v is the state of the system, beta defines the slope of change and s is the input, projected to zero dimensions by summing.
SerialOrderRecruiting	An Extension of the Serial Order Step, that allows a dynamic growth of serial order nodes and optionally a recruition of additional Groups toghether with the RecruitGroupsScript.
PatternMemory	A dynamics that relaxes to the input pattern as long as a learning input is active.
Image Processing
ConvertDepthToXyz	A step that converts depth data into 3D object data with scale and matrix center
EgoToBirdView	A step that converts camera input and distance information to a top-down representation.
GeometricImageTransform	Rescales and rotates and image around its center and then offsets it using geometric transformations.
SteerC2C3	Calculates the C2 and C3 component from steering filter responses.
SteerableFilter	Applies a steerable filter to an input.
SteeringAngle	Calculates the steering angle from C2/C3 responses.
SteeringEnergy	Calculates the steering energy from C2/C3 responses.
TransformPointCloud	A step that performs geometric operations on a point cloud, yielding the transformed point cloud and camera position.
Keypoints
DoubleStoppedCellResponses	Extracts the double-stopped cell responses from keypoint data.
KeypointExtractor	A step that extracts keypoint information.
KeypointLambdaMapping	A step that maps keypoint lambdas to other coordinate systems and vice versa.
KeypointLinesAndEdges	Extracts line- and edge information from a KeypointData object.
KeypointPatchRescaler	Rescales a local image patch according to a chosen keypoint size.
KeypointVisualization	Draws visualizations for a list of opencv keypoints onto an image.
Motion detection
CounterChangeCombination	The inputs have to be 3D toward and away matrices from a single edge filtered video or camera input. The first dimension of the input matrices is considered as the edge orientation dimension. The input matrices are shifted orthogonal the edge orientation by the size of the configurable shift parameter. Output matrices are two 3D matrices that reflect the combination of toward and away signals according to the counter-change rule. The first dimension is assigned to the motion direction. The two matrices represent the polarity if a dark pattern moves on bright ground (BtW: Black to White motion) or the other way around (WtB: White to Black motion)
MotionGradient	This class provides a cedar processing step implementing the OpenCV function calcMotionGradient.
Nao
NaoCamera	Nao's camera.
Neural Timer
ApproximateCoupling	bla
EquidistantMatrix	no description.
NeuralTimer	no description.
Object Recognition
FeatureStacks	This step builds localized histograms around keypoints. See Lomp et al. (2014) for details.
KeypointEdgeHistogramExtractor	This step builds localized histograms around keypoints using shape features.
ReceptiveFieldHistogram	Extracts histograms using localized receptive fields.
TopDownReconstruction	Reconstructs a shape based object representation.
ViewCombination	Subsamples discrete activation fields by summing up neighboring sampling points (1D only). This is intended for object recognition, where multiple nodes representing multiple object views may be summed to form a representation of a single object.
Oscillator
ApproximateCouplingVector	bla
ApproximateInput	bla
EquidistantRidge	no description.
HarmonicOscillator	A looped step that employs the harmonic Oscillator Formula pdotdot = -K(p-x)-B*pdot. K and B are automatically set to the critically damped solution.
Programming
BufferThief	This step can grab any buffer from a step in the same architecture and output the data in the buffer. Use with caution, as this may lead to issues due to unsafe/unlocked data.
Demultiplexer	Splits a vector (1xn or nx1 matrix) into individual scalars (1x1 matrices).
LabelString	Outputs a label and ordered list of labels for activation from a label field.
LabelWMAnnotator	Takes a yarp-friendly label WM representation and puts labels into an image.
MatrixThreadDecoupler	A looped step that makes a copy of its input matrix. This may help make threads more independent of each other.
Multiplexer	Joins several scalars (1x1 matrices) into a vector (nx1 matrix).
Sources
OpenGLTargetVisualisation	A step that visualizes a colored blob in the 3D Caren simulation
ImageProvider	A step that outputs an image from a directory based on an index specified via a parameter.
KinectReader	A source that reads a Kinect camera, yielding images and depth images, as well as a RGBA point cloud.
LinearSpatialPattern	A linear spatial pattern
SpatialPattern	Outputs a matrix of synaptic weights that correspond to a spatial relational template (e.g., "to the left of").

Group templates included in the plugin

group template	description
edge histogram rotation	Template forward rotating localized edge histograms within the CF thesis architecture.
edge rotation estimation	Mathing of rotation based on localized edge histograms.
histogram label matching	Mathing for labels based on localized histograms.
histogram shift estimation	Mathing of shift based on localized histograms.
localized histogram extraction	Template for extracting localized histograms within the CF thesis architecture.
localized histogram forward shift	Template forward shifting localized histograms within the CF thesis architecture.
steering filter bank	Applies a bank of steering filters and calculates energy and angle for each pixel.

Devices included in the plugin

Kernels included in the plugin

kernel	description
MotionKernel	This is a general version of the gauss kernel class. A global inhibition term can be added to each dimension separately. There is an amplitude parameter which scales the gaussian local excitation of each dimension but not the global inhibition term of this dimension. The benefit of this class in contrast to the usual Gauss kernel is that it is possible to introduce global excitation along only one dimension. This is, e.g., important for motion detection where you want to detect a single direction selective motion percept (global inhibition for motion direction) but for multiple objects (no global inhibition in the space dimension).
SteerableKernel	A collection of steerable kernels that can be used to quickly calculate edge orientations and strengths.

Data structures included in the plugin

data structure	description
KeypointData	A data structure that holds a vislab::keypoints::KPData object.
KeypointListData	A std::vector of (open)cv::Keypoints.
RGBAPointCloudData	A data structure that holds Kinect point cloud data.
StringData	A data structure that holds a string.
AsyncGrabber	A data structure that holds refreshed Kinect image data.

Plots included in the plugin

plot	description
StringPlot	A plot that displays a string (cedar::aux::StringData).

External libraries used by the plugin

naoqi

TODO describe how to get, build and include the naoqi.

vislab_toolbox

TODO describe how to get, build and include the vision lab toolbox.