Example Scene Classification

From BoofCV
Revision as of 16:14, 17 January 2022 by Peter (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search

Scene classification is the problem where you are presented with an image and you need to classify it as belonging to a known set. The example below demonstrates how to perform scene classification using dense SURF features and a nearest neighbour classifier. See code documentation for a more detailed discussion.

Example Code:


  • Scene Classification
  • Dense Image Features
  • Clustering
  • k-NN classifier

Related Examples:

Example Code

 * <p>
 * Example of how to train a K-NN bow-of-word classifier for scene recognition. The resulting classifier
 * produces results which are correct 52.2% of the time. To provide a point of comparison, randomly selecting
 * a scene is about 6.7% accurate, SVM One vs One RBF classifier can produce accuracy of around 74% and
 * other people using different techniques claim to have achieved around 85% accurate with more advanced
 * techniques.
 * </p>
 * Training Steps:
 * <ol>
 * <li>Compute dense SURF features across the training data set.</li>
 * <li>Cluster using k-means to create works.</li>
 * <li>For each image compute the histogram of words found in the image</li>
 * <li>Save word histograms and image scene labels in a classifier</li>
 * </ol>
 * Testing Steps:
 * <ol>
 * <li>For each image in the testing data set compute its histogram</li>
 * <li>Look up the k-nearest-neighbors for that histogram</li>
 * <li>Classify an image by by selecting the scene type with the most neighbors</li>
 * </ol>
 * <p>NOTE: Scene recognition is still very much a work in progress in BoofCV and the code is likely to be
 * significantly modified in the future.</p>
 * @author Peter Abeles
public class ExampleClassifySceneKnn extends LearnSceneFromFiles {

	// Tuning parameters
	public static int NUMBER_OF_WORDS = 100;
	public static boolean HISTOGRAM_HARD = true;
	public static int NUM_NEIGHBORS = 10;
	public static int MAX_KNN_ITERATIONS = 100;

	// Files intermediate results are stored in
	public static final String CLUSTER_FILE_NAME = "clusters.obj";
	public static final String HISTOGRAM_FILE_NAME = "histograms.obj";

	// Algorithms
	ClusterVisualWords cluster;
	DescribeImageDense<GrayU8, TupleDesc_F64> describeImage;
	NearestNeighbor<HistogramScene> nn;

	ClassifierKNearestNeighborsBow<GrayU8, TupleDesc_F64> classifier;

	public ExampleClassifySceneKnn( final DescribeImageDense<GrayU8, TupleDesc_F64> describeImage,
									ComputeClusters<double[]> clusterer,
									NearestNeighbor<HistogramScene> nn ) {
		this.describeImage = describeImage;
		this.cluster = new ClusterVisualWords(clusterer, 0xFEEDBEEF);
		this.nn = nn;

	 * Process all the data in the training data set to learn the classifications. See code for details.
	public void learnAndSave() {
		System.out.println("======== Learning Classifier");

		// Either load pre-computed words or compute the words from the training images
		AssignCluster<double[]> assignment;
		if (new File(CLUSTER_FILE_NAME).exists()) {
			assignment = UtilIO.load(CLUSTER_FILE_NAME);
		} else {
			System.out.println(" Computing clusters");
			assignment = computeClusters();

		// Use these clusters to assign features to words
		var featuresToHistogram = new FeatureToWordHistogram_F64(assignment, HISTOGRAM_HARD);

		// Storage for the work histogram in each image in the training set and their label
		List<HistogramScene> memory;

		if (!new File(HISTOGRAM_FILE_NAME).exists()) {
			System.out.println(" computing histograms");
			memory = computeHistograms(featuresToHistogram);
			UtilIO.save(memory, HISTOGRAM_FILE_NAME);

	 * Extract dense features across the training set. Then clusters are found within those features.
	private AssignCluster<double[]> computeClusters() {
		System.out.println("Image Features");

		// Compute features in the training image set
		var features = new ArrayList<TupleDesc_F64>();
		for (String scene : train.keySet()) {
			List<String> imagePaths = train.get(scene);
			System.out.println("   " + scene);

			for (String path : imagePaths) {
				GrayU8 image = UtilImageIO.loadImage(path, GrayU8.class);

				// the descriptions will get recycled on the next call, so create a copy
				for (TupleDesc_F64 d : describeImage.getDescriptions()) {
		// add the features to the overall list which the clusters will be found inside of
		for (int i = 0; i < features.size(); i++) {

		// Find the clusters. This can take a bit

		UtilIO.save(cluster.getAssignment(), CLUSTER_FILE_NAME);

		return cluster.getAssignment();

	public void loadAndCreateClassifier() {
		// load results from a file
		List<HistogramScene> memory = UtilIO.load(HISTOGRAM_FILE_NAME);
		AssignCluster<double[]> assignment = UtilIO.load(CLUSTER_FILE_NAME);

		var featuresToHistogram = new FeatureToWordHistogram_F64(assignment, HISTOGRAM_HARD);

		// Provide the training results to K-NN and it will preprocess these results for quick lookup later on
		// Can use this classifier with saved results and avoid the

		classifier = new ClassifierKNearestNeighborsBow<>(nn, describeImage, featuresToHistogram);
		classifier.setClassificationData(memory, getScenes().size());

	 * For all the images in the training data set it computes a {@link HistogramScene}. That data structure
	 * contains the word histogram and the scene that the histogram belongs to.
	private List<HistogramScene> computeHistograms( FeatureToWordHistogram_F64 featuresToHistogram ) {

		List<String> scenes = getScenes();

		List<HistogramScene> memory;// Processed results which will be passed into the k-NN algorithm
		memory = new ArrayList<>();

		for (int sceneIndex = 0; sceneIndex < scenes.size(); sceneIndex++) {
			String scene = scenes.get(sceneIndex);
			System.out.println("   " + scene);
			List<String> imagePaths = train.get(scene);

			for (String path : imagePaths) {
				GrayU8 image = UtilImageIO.loadImage(path, GrayU8.class);

				// reset before processing a new image
				for (TupleDesc_F64 d : describeImage.getDescriptions()) {

				// The histogram is already normalized so that it sums up to 1. This provides invariance
				// against the overall number of features changing.
				double[] histogram = featuresToHistogram.getHistogram();

				// Create the data structure used by the KNN classifier
				HistogramScene imageHist = new HistogramScene(NUMBER_OF_WORDS);
				imageHist.type = sceneIndex;

		return memory;

	protected int classify( String path ) {
		GrayU8 image = UtilImageIO.loadImage(path, GrayU8.class);

		return classifier.classify(image);

	public static void main( String[] args ) {

		var surfFast = new ConfigDenseSurfFast(new DenseSampling(8, 8));
//		ConfigDenseSurfStable surfStable = new ConfigDenseSurfStable(new DenseSampling(8,8));
//		ConfigDenseSift sift = new ConfigDenseSift(new DenseSampling(6,6));
//		ConfigDenseHoG hog = new ConfigDenseHoG();

		DescribeImageDense<GrayU8, TupleDesc_F64> desc =
				FactoryDescribeImageDense.surfFast(surfFast, GrayU8.class);
//				FactoryDescribeImageDense.surfStable(surfStable, GrayU8.class);
//				FactoryDescribeImageDense.sift(sift, GrayU8.class);
//				FactoryDescribeImageDense.hog(hog, ImageType.single(GrayU8.class));

		var configKMeans = new ConfigKMeans();
		configKMeans.maxIterations = MAX_KNN_ITERATIONS;
		configKMeans.reseedAfterIterations = 20;
		ComputeClusters<double[]> clusterer = FactoryClustering.kMeans_MT(
				configKMeans, desc.createDescription().size(), 200, double[].class);
		// The _MT tells it to use the threaded version. This can run MUCH faster.

		int pointDof = desc.createDescription().size();
		NearestNeighbor<HistogramScene> nn = FactoryNearestNeighbor.exhaustive(new KdTreeHistogramScene_F64(pointDof));
		ExampleClassifySceneKnn example = new ExampleClassifySceneKnn(desc, clusterer, nn);

		var trainingDir = new File(UtilIO.pathExample("learning/scene/train"));
		var testingDir = new File(UtilIO.pathExample("learning/scene/test"));

		if (!trainingDir.exists() || !testingDir.exists()) {
			String addressSrc = "http://boofcv.org/notwiki/largefiles/bow_data_v001.zip";
			File dst = new File(trainingDir.getParentFile(), "bow_data_v001.zip");
			try {
				DeepBoofDataBaseOps.download(addressSrc, dst);
				DeepBoofDataBaseOps.decompressZip(dst, dst.getParentFile(), true);
				System.out.println("Download complete!");
			} catch (IOException e) {
				throw new UncheckedIOException(e);
		} else {
			System.out.println("Delete and download again if there are file not found errors");
			System.out.println("   " + trainingDir);
			System.out.println("   " + testingDir);

		example.loadSets(trainingDir, null, testingDir);
		// train the classifier
		// now load it for evaluation purposes from the files

		// test the classifier on the test set
		Confusion confusion = example.evaluateTest();
		System.out.println("Accuracy = " + confusion.computeAccuracy());

		// Show confusion matrix
		// Not the best coloration scheme... perfect = red diagonal and blue elsewhere.
		ShowImages.showWindow(new ConfusionMatrixPanel(
				confusion.getMatrix(), example.getScenes(), 400, true), "Confusion Matrix", true);

		// For SIFT descriptor the accuracy is          54.0%
		// For  "fast"  SURF descriptor the accuracy is 52.2%
		// For "stable" SURF descriptor the accuracy is 49.4%
		// For HOG                                      53.3%

		// SURF results are interesting. "Stable" is significantly better than "fast"!
		// One explanation is that the descriptor for "fast" samples a smaller region than "stable", by a
		// couple of pixels at scale of 1. Thus there is less overlap between the features.

		// Reducing the size of "stable" to 0.95 does slightly improve performance to 50.5%, can't scale it down
		// much more without performance going down