(038) previous ~ index ~ next
To: TDT <tdt-distrib@ldc.upenn.edu>
From: George Doddington <doddington@nist.gov>
Subject: Re: a note on the TDT segmentation cost function
Date: Fri, 05 Mar 1999 13:20:12 -0500
I would like to follow up my response with a few more comments:
I have the sense that there is a subtle implication that the "optimum" segmentation algorithm, as determined by the segmentation evaluation, should also be optimum for use in tracking/detection. This is not the case, as Dragon has pointed out. Source characteristics and evaluation conditions may dictate the use of different segmentation parameter settings. In fact, as Dragon suggests, it may well be desirable to use different parameter values for different sources, according to expected story length statistics. Separate work on (and evaluation of) segmentation, however, is still an important research focus.
While cost-based evaluation provides a single-number objective target for research evaluation, it is not very intuitive, as Dragon has pointed out. It is, however, derived from a reasonable application model. And the error probabilities that underlie it DO supply good intuitive characterization of performance, given the window method we use to calculate error probabilities. Using a window of T seconds, for example, supports a very nice gray-scale intuitive impression of performance (e.g., Pmiss = t/T for a deviation of t seconds, and Pfa = T/t for a false alarm period of t seconds).
--
George Doddington at NIST: doddington@nist.gov or 301/975-3261
(038) previous ~ index ~ next
Last updated Thu May 13 09:28:18 1999