Several sets of experimental tests were conducted to aid in development, to test the sensitivity of the algorithms, and determine accuracy and false alarm rate. By way of example, Figure 8 shows a container image which the algorithm correctly identified as non-empty. Figure 9 shows an image that was a false positive – the algorithm identified the container as non-empty when in fact it was empty.
We tested both algorithms with several thousand test images. The rule-based algorithm achieved an accuracy of 97.2% and false negative rate of 0.4%, while the statistical algorithm reached an accuracy of 96.5% and false negative rate of 2.3%. The false negative rates are considered most important by initial customers, so we are working hard to reduce them. There are still many fertile areas for improving performance and we are confident the false negative rates will continue to drop.
We tested both algorithms with several thousand test images. The rule-based algorithm achieved an accuracy of 97.2% and false negative rate of 0.4%, while the statistical algorithm reached an accuracy of 96.5% and false negative rate of 2.3%. The false negative rates are considered most important by initial customers, so we are working hard to reduce them. There are still many fertile areas for improving performance and we are confident the false negative rates will continue to drop.
No comments:
Post a Comment