EIGEN WINDOWS Sample Clauses

EIGEN WINDOWS. The Eigen window approach is almost the same as the general Eigen space technique. Using Eigen windows, widows (often found from point of interest operators) are substituted for whole images. The process then continues as discussed. The Eigen window work undertaken in this project was initially based on other work by [13] [14] and [16]. Some windows may be rejected based on the low amount of variation they capture, found by taking the average change in pixel intensity value throughout the window, then using a suitable threshold value. Once the projection data is obtained, training points are projected to create the Eigen space. Global Eigen point culling is then performed so that the optimum Eigen space for classification is created [13]. Figure 6.4 – Eigen Window Patch Matching: Green Patches Represent Correctly Classified Patch Sections, Red Represent Incorrect Points

EIGEN WINDOWS. Eigen window based patch­matching techniques, are shown to be usable and robust for classification tasks such as these. The classification rates are almost all in the 80 – 90% correct classification range, which although not as good as some techniques published, is a reasonable result considering the low camera resolution and complexity of the task. The top scoring Eigen window based technique, which used scale invariant patches for matching, robustly and continually produced classification scores in the high 80% and low 90% range. Speed is a concern when using this technique and considerable optimisation may have to be undertaken before this technique as implemented can be used in a frame rate application, especially if a large patch size is to be used. This technique deals with the problems of image noise, which can be a problem for affine invariant techniques, as the system for resolving the characteristic scale uses very robust criteria. Surprisingly, the next highest scoring Eigen space based technique was the unchanged (simple Xxxxxx point) Eigen window. This patch had no tolerance to the different distortions present within the project, and so was not expected to perform well. The main reason for its high score rate in this instance is the similarity between the testing and training sequence. Similar object views were available in both, so the scale and rotation difference did not overly effect the classification. Had a different training or testing sequence been used, this technique may have performed less admirably. Before testing, the SIFT adaptations to the Xxxxxx based Eigen window technique was expected to outperform both the unchanged Eigen window and Whitening normalised Eigen window technique. After testing, it became apparent that although this technique is normally very robust, the reduced clarity of the cameras used and non­linear lighting effects adversely affected the histogram orientation assignment process upon which this technique relies. This is also apparent in techniques such as MSER and salient region detection. Although many improvement criteria exist, the criteria used here was included to improve the quality of the patches used, while reducing as many non­informative image patches from the matching space as possible. Although the improvement criteria did reduce the overall number of patches within the matching space, thus delivering a large decrease in time requirements per patch classification, many of the patches removed we...