Introduction

Direct Graphical Models (DGM)

is a C++ dynamic link library implementing various tasks in probabilistic graphical models with pairwise dependencies as well as complete (dense) graphs. The library aims to be used for the Markov- and Conditional Random Fields (MRF / CRF), Markov Chains, Bayesian Networks, etc. DGM library consists of three modules:

• Main DGM Module, which includes a variety of methods for the tasks:
  • Training
  • Graph Building
  • Inference / Decode
  • Parameter Estimation
  • Evaluation
• Feature extraction FEX Module, which allows for extracting the main data features, used nowadays in image classification.
  • Local Features Extraction for pixel-level features.
  • Global Features Extraction for image-level features.
• Visualization VIS Module, wich allows for visualizing the used data and features as well as intermediate and final classification results.

These tasks are optimized for speed, i.e. high-efficient calculations. The code is written entirely in C++ and can be compiled with Microsoft Visual C++.

Methods Overview

Training

DGM implements the following training methods:

Nodes | Unary Potentials

• NaiveBayes: Naive Bayes training DirectGraphicalModels::CTrainNodeBayes
• GMM: Gaussian Mixture Model training (Sequential GMM Training Algorithm) DirectGraphicalModels::CTrainNodeGMM
• CvGMM: OpenCV Gaussian Mixture Model training DirectGraphicalModels::CTrainNodeCvGMM
• KNN: k-Nearest Neighbors training DirectGraphicalModels::CTrainNodeKNN
• CvKNN OpenCV k-Nearest Neighbors training DirectGraphicalModels::CTrainNodeCvKNN
• CvRF: OpenCV Random Forest training DirectGraphicalModels::CTrainNodeCvRF
• MsRF: Microsoft Research Random Forest training DirectGraphicalModels::CTrainNodeMsRF
• CvANN: OpenCV Artificial Neural Network training DirectGraphicalModels::CTrainNodeCvANN
• CvSVM: OpenCV Support Vector Machine training DirectGraphicalModels::CTrainNodeCvSVM

The corresponding classes are CTrainNode* (where * is the name of the method above). The difference between these methods is described at forum: Training of a Random Model.

Edges | Pairwise Potentials

• Potts: Train- & Test-data-independent Potts model DirectGraphicalModels::CTrainEdgePotts
• PottsCS: Train-data-independent contrast-sensitive Potts model DirectGraphicalModels::CTrainEdgePottsCS
• Prior: Contrast-sensitive Potts model with prior probability DirectGraphicalModels::CTrainEdgePrior
• Concat: Concatenated training DirectGraphicalModels::CTrainEdgeConcat

The corresponding classes are CTrainEdge* (where * is the name of the method above).

Inference / Decode

DGM implements the following inference and decoding methods:

Inference

• Exact: Exact inferece for small graphs with an exhaustive search DirectGraphicalModels::CInferExact
• Chain: Exact inferece for Markov chains (chain-structured graphs) DirectGraphicalModels::CInferChain
• Tree: Exact inferece for undirected graphs without loops (tree-structured graphs) DirectGraphicalModels::CInferTree
• LBP: Approximate inference based on the Loopy Belief Propagation (sum-product message-passing) algorithm DirectGraphicalModels::CInferLBP
• TRW: Approximate inference based on the (Convergent Tree-Reweighted) (max-sum message-passing) algorithm DirectGraphicalModels::CInferTRW
• Viterbi: Approximate inference based on Viterbi (max-sum message-passing) algorithm DirectGraphicalModels::CInferViterbi
• Dense: Efficient inference for dense CRFs with Gaussian edge potentials (paper) DirectGraphicalModels::CInferDense

The corresponding classes are CInfer* (where * is the name of the method above).

All of the inference classes may be also used for approximate decoding via function DirectGraphicalModels::CInfer::decode()

Decoding

• Exact: Exact decoding for small graphs with an exhaustive search DirectGraphicalModels::CDecodeExact

The corresponding classes are CDecode* (where * is the name of the method above).

Parameter Estimation

DGM implements the following parameter estimation method:

• Powell: Powell search method DirectGraphicalModels::CPowell

Sampling

DGM implements the following sampling method:

• Gauss: Sampling from the multivariate gaussian distribution DirectGraphicalModels::CKDGauss::getSample()

Feature Extraction

Please refer to the FEX Module documentation

Visualization

Please refer to the VIS Module documentation

Quick Links

• Installation
• Tutorials
• User Q&A forum
• Report a bug

Author

Sergey G. Kosov, serge.nosp@m.y.ko.nosp@m.sov@p.nosp@m.roje.nosp@m.ct-10.nosp@m..de