drwnVision

Class List for drwnVision

• drwnColourHistogram : Specialized histogram for quantized 3-channel colour values (e.g., RGB).
• drwnFilterBankResponse : Holds the results of running an image through a bank of filters and allows for computation of features over rectangular regions.
• drwnGrabCutInstance : Implements the grabCut algorithm of Rother et al., SIGGRAPH 2004 for figure/ground segmentation.
• drwnHOGFeatures : Encapsulates histogram-of-gradient (HOG) feature computation.
• drwnImageCache : Caches images in memory up to a maximum number of images or memory limit.
• drwnImageInPainter : Performs exemplar-based image inpainting.
• drwnImagePyramidCache : Caches image pyramids in main memory up to a maximum number of images or memory limit.
• drwnMaskedPatchMatch : Implements the basic PatchMatch algorithm of Barnes et al., SIGGRAPH 2009 on masked images.
• drwnMultiSegRegionDefinitions : Provides a mechanism for mapping region IDs to colours and class names. Can be initialized from an XML configuration file or programmatically for a number of standard datasets.
• drwnMultiSegConfig : Manages configuration settings for multiple image segmentation.
• drwnMultiSegVis : Visualization routines for multi-class image segmentation.
• drwnObject : Encapsulates a 2D object in an image for object detection.
• drwnObjectList : List of objects for the same image (see drwnObject)
• drwnObjectSequence : Sequence of images, each with a list of objects (see drwnObjectList)
• drwnOpenCVUtils : Utilities for working with OpenCV.
• drwnPatchMatchEdge : Represents an edge in the drwnPatchMatchGraph.
• drwnPatchMatchGraph : Each image maintains a W-by-H-by-K array of match records referencing the (approximate) best K matches to other images. Image filename rather than the images themselves are stored. Duplicate filenames (images) are not allowed.
• drwnPatchMatchGraphLearner : Learns a PatchMatchGraph by iteratively performing search moves over the space of matches.
• drwnPatchMatchImagePyramid : Record of patch matches for mutliple levels each image.
• drwnPatchMatchImageRecord : Records matches for one level in an image pyramid.
• drwnPatchMatchNode : Represents a node in the drwnPatchMatchGraph.
• drwnPatchMatchRepaint :
• drwnPartsModel : Interface for implementing a part-based constellation model (i.e., pictorial structures model) for object detection.
• drwnPixelNeighbourContrasts : Convenience class for holding pixel contrast weights.
• drwnPixelSegCRFInference : Alpha-expansion inference for a pixel-level CRF model with unary, contrast-dependent pairwise, and custom higher-order terms.
• drwnPixelSegModel : Implements a pixel-level CRF model for multi-class image segmentation (pixel labeling).
• drwnRobustPottsCRFInference : and where 0.0 < < 0.5
• drwnSegImageInstance : Encapsulates a single instance of an image for multi-class pixel labeling problems (i.e., image segmentation).
• drwnSegImagePixelFeatures : Interface for generating per-pixel features for a drwnSegImageInstance object.
• drwnSegImageRegionFeatures : Interface for generating per-region (or per-superpixel) features for a drwnSegImageInstance object. The superpixel data member of the drwnSegImageInstance object must be populated.
• drwnSuperpixelContainer : Holds multiple oversegmentations for a given image.
• drwnTemplateMatcher : Utility class for computing multiple template matches.
• drwnTextonFilterBank : Implements a 17-dimensional filter bank.
• drwnVisionUtils : Computer vision utility functions (not dependent on OpenCV functions, but may use OpenCV data structures).
• drwnWeightedRobustPottsCRFInference : and where 0.0 < < 0.5, W_c = w_i

drwnVision API Examples

The following code shows an example of learning and then executing a pixel classifier based on RGB features. For more sophisticated pixel labeling see drwnSegImageInstance and drwnPixelSegModel.

// load image and labels
cv::Mat img = cv::imread("image.png", CV_LOAD_IMAGE_COLOR);
MatrixXi labels;
drwnLoadPixelLabels(labels, "labels.txt");

// build a training dataset
drwnClassifierDataset dataset;
dataset.reserve(img.size());

for (int y = 0; y < img.rows; y++) {
    for (int x = 0; x < img.cols; x++) {
        if (labels(y, x) < 0) continue;

        vector<double> features(3);
        features[0] = img.at<Vec3b>(y, x)[0];
        features[1] = img.at<Vec3b>(y, x)[1];
        features[2] = img.at<Vec3b>(y, x)[2];

        dataset.append(features, labels(y, x));
    }
}
DRWN_LOG_MESSAGE("dataset contains " << dataset.size() << " training samples");

dataset.subSample(2, true);
DRWN_LOG_MESSAGE("subsampled to " << dataset.size() << " training samples");

// train a decision tree classifier to depth of 10
const int nClasses = dataset.maxTarget() + 1;
drwnDecisionTree::MAX_DEPTH = 10;
drwnDecisionTree classifier(dataset.numFeatures(), nClasses);
classifier.train(dataset);

// display classification results (on training set)
drwnConfusionMatrix results(nClasses);
results.accumulate(dataset, &classifier);
results.printCounts();

The following code shows an example of drawing bounding boxes for a list of objects.

// load objects
drwnObjectList objects;
objects.read(OBJECTS_FILENAME);

// load image
cv::Mat img = cv::imread(IMAGE_FILENAME, CV_LOAD_IMAGE_COLOR);

// draw objects
for (drwnObjectList::const_iterator it = objects.begin(); it != objects.end(); it++) {
    drwnDrawBoundingBox(img, it->extent, CV_RGB(255, 0, 0), CV_RGB(0, 0, 0), 2);
}

// show image with objects
drwnShowDebuggingImage(img, "objects", true);