assignment criteria <strong>for</strong> the FR <strong>algorithm</strong>. In this modified <strong>for</strong>mulation, the FR<strong>algorithm</strong> extracts from each shot the same number of <strong>key</strong> <strong>frame</strong>s computed bythe CP <strong>algorithm</strong>. The same criterion was applied to the ATS <strong>algorithm</strong>.Table 5, Table 6, <strong>an</strong>d Table 7 show the results of the different <strong>algorithm</strong>s on thesix <strong>video</strong> sequences. Table 5 shows the Compression Ratio (CRatio) values <strong>an</strong>dsome statistics. KFPerShot <strong>in</strong>dicates the average number of <strong>key</strong> <strong>frame</strong>s <strong>for</strong> eachshot: nUsedShots <strong>an</strong>d FrPerShots <strong>in</strong>dicate the number of type A shots <strong>an</strong>d theaverage number of <strong>frame</strong>s per shots respectively. From Table 5 we c<strong>an</strong> see that <strong>for</strong>the “eeopen”,”nw<strong>an</strong>w1” <strong>an</strong>d “bugsbunny” <strong>video</strong>s some type B shots have beenremoved by our shot detection <strong>algorithm</strong>. Our <strong>key</strong> <strong>frame</strong> selection <strong>algorithm</strong> isable to extract more th<strong>an</strong> one <strong>frame</strong> per type A shot as is the PME <strong>algorithm</strong> (thatthe number of <strong>key</strong> <strong>frame</strong>s of the FR, ATS <strong>an</strong>d SRDI <strong>algorithm</strong>s depends by theCP <strong>algorithm</strong>). Videos with high dynamics, <strong>in</strong> fact, are assigned more <strong>key</strong> <strong>frame</strong>sper shot th<strong>an</strong> those with little motion. This c<strong>an</strong> be seen <strong>in</strong> case of the“basketball2”, “football” <strong>an</strong>d “news” <strong>video</strong>s. For the “bugsbunny” <strong>video</strong>, althoughvery few shots exhibit high dynamics, the PME <strong>algorithm</strong> extracts nearly doublethe number of <strong>key</strong> <strong>frame</strong>s th<strong>an</strong> the CP <strong>algorithm</strong>.Table 6 gives the Fidelity measure results with the m<strong>in</strong>imum <strong>an</strong>d st<strong>an</strong>darddeviation of the measure computed on both the histogram <strong>an</strong>d the HWDdescriptors. Table 7 presents the results of the Shot Reconstruction Degreemeasure. They show that the average Fidelity <strong>an</strong>d SRD measures computed on thehistogram (HST) are less variable th<strong>an</strong> those computed on the HWD. This seemsto <strong>in</strong>dicate that <strong>frame</strong> differences tend to be less dist<strong>in</strong>guishable when us<strong>in</strong>g thecolor histogram alone. It is also <strong>in</strong>terest<strong>in</strong>g to note that the summaries of the“basketball2” <strong>video</strong> are clearly unacceptable if we evaluate them with the Fidelitymeasure computed on the histogram, while they appear fairly acceptable ifevaluated with the HWD.To judge the per<strong>for</strong>m<strong>an</strong>ce of our <strong>algorithm</strong> with respect to the other five<strong>algorithm</strong>s, it is useful to express the results as a measure of relative improvement(∆Q) us<strong>in</strong>g the follow<strong>in</strong>g <strong>for</strong>mula:( Measure_Alg( CP) − Measure_Alg( X ))Measure_Alg( X )∆Q( X ) =(21)where Measure_Alg corresponds to the Fidelity <strong>an</strong>d the SRD measure, <strong>an</strong>d wesubstitute X with FR, ATS, SRDI, PME <strong>an</strong>d MP <strong>in</strong> turn.Table 8 details <strong>for</strong> each <strong>video</strong> the relative edge of the per<strong>for</strong>m<strong>an</strong>ce of our<strong>algorithm</strong> over that of other five <strong>algorithm</strong>s. It c<strong>an</strong> be seen that when the numberof <strong>key</strong> <strong>frame</strong>s per shot is small as <strong>in</strong> the case of the “eeopen” <strong>an</strong>d “bugsbunny”<strong>video</strong>s (1,136 <strong>an</strong>d 1,145 <strong>key</strong> <strong>frame</strong>s per shot respectively) the differences betweenthe CP, FR, ATS <strong>an</strong>d MP <strong>algorithm</strong>s are slight. Exceptions are the SRDI<strong>algorithm</strong>, about 5% worse th<strong>an</strong> the CP <strong>algorithm</strong>, <strong>an</strong>d the PME <strong>algorithm</strong> <strong>in</strong> the“bugsbunny” <strong>video</strong>, about 4% better th<strong>an</strong> the CP <strong>algorithm</strong> (note that the PME<strong>algorithm</strong> extracts 182 <strong>key</strong> <strong>frame</strong>s <strong>an</strong>d the CP <strong>algorithm</strong> 95). For the other <strong>video</strong>s,the PME <strong>algorithm</strong> shows, <strong>in</strong>stead, the worst results (exclud<strong>in</strong>g the MP<strong>algorithm</strong>). As the number of <strong>key</strong> <strong>frame</strong>s per shot <strong>in</strong>creases, the gap is moremarked. With the exception of the “basketball2” <strong>video</strong>, the per<strong>for</strong>m<strong>an</strong>ces of theFR <strong>algorithm</strong> <strong>an</strong>d the CP <strong>algorithm</strong> do not exhibit large differences; only <strong>in</strong> the“news” <strong>video</strong> does the FR <strong>algorithm</strong> outper<strong>for</strong>m (slightly) the CP <strong>algorithm</strong>. Themost <strong>in</strong>terest<strong>in</strong>g experiment is the one regard<strong>in</strong>g the “basketball2” <strong>video</strong>. The highdynamics <strong>an</strong>d the length of the shot, cause greatly diverg<strong>in</strong>g results. The20
per<strong>for</strong>m<strong>an</strong>ce of the SRDI <strong>algorithm</strong> exhibits a variable behavior depend<strong>in</strong>g on themeasure employed, while the MP <strong>algorithm</strong> shows the worse results.Table 9 lists the ∆Q measure of relative improvement computed as thepercentage average <strong>for</strong> all five <strong>video</strong>s. Overall our <strong>algorithm</strong> outper<strong>for</strong>ms theother five <strong>algorithm</strong>s. Its adv<strong>an</strong>tages over the FR <strong>algorithm</strong> is negligible, but wemust remember that the FR <strong>algorithm</strong> uses <strong>an</strong> optimization strategy to allocate the<strong>key</strong> <strong>frame</strong>s with<strong>in</strong> a shot, <strong>an</strong>d requires that the number of <strong>key</strong> <strong>frame</strong>s must begiven a priori. For the ATS <strong>an</strong>d SRDI <strong>algorithm</strong>s, the gap is more noticeable, <strong>an</strong>ds<strong>in</strong>ce the number of <strong>key</strong> <strong>frame</strong>s extracted are the same, the results depend only onthe selection strategies adopted by the <strong>algorithm</strong>s. The poor per<strong>for</strong>m<strong>an</strong>ce of theMP <strong>algorithm</strong> is, of course, to be expected s<strong>in</strong>ce it extracts only one <strong>key</strong> <strong>frame</strong>sfrom each shot.More <strong>in</strong>terest<strong>in</strong>g is the per<strong>for</strong>m<strong>an</strong>ce of the PME <strong>algorithm</strong>, which extracts <strong>key</strong><strong>frame</strong>s <strong>in</strong> a totally automatic way based on motion vectors. The results of the SRDmeasure show that the CP <strong>algorithm</strong> outper<strong>for</strong>ms the PME <strong>algorithm</strong> by about10%. The <strong>key</strong> <strong>frame</strong>s selected only by motion do not seem to be successful <strong>in</strong>visually represent<strong>in</strong>g the content of the <strong>video</strong>.21