Expert Object Recognition in Video在视频专家标识别.pptVIP

Expert Object Recognition in Video Matt McEuen The EOR Pathway Early vision (feature extraction) Categorization Exemplar matching Expert Object Recognition Early Vision: Edge Detection Gabor filters Three filter sizes Four orientations Even and odd Early Vision: Edge Detection Early Vision: Line Detection Non-accidental structural properties collinearity parallelism symmetry Hough transform Categorization Allows a unique subspace for each category K-Means Categorization Exemplar Matching Principal Component Analysis (PCA) Based on covariance Visual memory reconstruction PCA Calculate covariance matrix of the samples Get the eigenvectors of the covariance matrix Choose which eigenvectors to keep Transform the data with the resulting matrix Exemplar Matching VENUS Biologically inspired Habituation Low-level features Knowledge and Hierarchical Learning Architecture Benefits of EOR for video General-purpose Segmentation: Attention window Associative memory in VENUS Problems with EOR for video Why learning? Problems with EOR for video Hard training / testing distinction Lots of processing The parameter k Data Data VEOR architecture VEOR architecture VEOR architecture VEOR architecture VEOR architecture Foreground segmentation Object tracking Feature extraction Clustering: goals Automatically determine k Facilitate learning ... efficiently Clustering: solutions Reuse of cluster centroids Clustering: solutions Reuse of cluster centroids Cluster growing and splitting Clustering: solutions Reuse of cluster centroids Cluster growing and splitting Incremental clustering Clustering: Fuzzy K-means Clustering Exemplar matching PCA Dirty flags Output: best match distance Membership subsystem Associative memory Multi-class [0,1] hypothesis One exemplar match per image One membership hypothesis per tracked object Three kinds of exemplar matching Matching to a training image Matching to a different learned object Matching to the same learned object Exemplar match: training i