Semester: 10th Semester, Master Thesis Title
Implementing Test Results in the Smile Assessment Algorithm
Download 0.97 Mb.
|
- Navigate this page:
- 8.3.Test phase two
- 8.3.1.Test Results – Test Subjects
- 8.3.2.Test Results – Computer
8.2.Implementing Test Results in the Smile Assessment AlgorithmTable - Test Phase One - Test Subjects, average rating - Graph Table - Test Phase One - Computer Results - Graph Figure - Detection area too small, chin - Picture 20 Figure - Detection area too large, right side - Picture 16 Figure - Detection error, chin - Picture 15 Figure - Detection area too small – Picture 9 The following graphs show the comparison between the smile estimation by the computer and the smile ratings given by the test groups in test phase one. The value of the x-axis corresponds to picture 1, 2 and so forth. The value of the y-axis represents the rating scale (1-10). The computer estimation is based on the distance between the centre of point A and B to point D. The distance is an expression of the level of smile in each picture calculated by the computer. In direct comparison from the graphs, the computer estimates of the smile are correct for Picture 1,3, 4, 5, 6, 10, 12, 13, 14 and 18. But for Pictures 2, 7, 8, 9, 11, 15, 16, 17, 19, 20 they are incorrect. In the pictures from test phase one, the algorithm calculated a 50% correct estimation. The peaks and valleys from each graph were compared to one another without regard to the intensity (i.e. Picture 4 has a value of 9.2 from test subjects whereas the computer gave it a rating of 1.97). Of interest was if the valleys and peaks corresponded to the same valleys and peaks in each graph. The pictures that were given erroneous ratings by the computer will be analysed in the following to establish in what way the algorithm failed to rate correctly. Each picture was re-estimated by the algorithm from this re-estimation, the following pictures (Figure 18, 19, 20 & 21) were removed due to erroneous detections of the facial region and facial features.     The pictures shown in, Figure 18, 19, 20 & 21 were given an erroneous smile rating by the computer due to the display of the specific smile. The smile in picture 2 is neutral (rating by test participants: 4.1) the computer does not register either a big/small smile and therefore provides a low rating. The smile in picture 7 indicates a big smile to the computer due to the placement of the corners of the mouth, but test participants rate the smile as only 6.24, slightly over neutral. Picture 8 displays a happy child (rating by test participants: 7.66) but the computer sees it as a neutral smile due to the corners of the mouth not being higher up. Test participants rated picture 11 as a 2.56 but due to the open mouth the computer finds the smile to be of higher value than test participants. Picture 17 was rated 4.22 but the computer saw a low smile due to the placement of the corners of the mouth. Due to these findings the following test phase two will have to be altered in order to answer the questions proposed in the problem statement at the beginning of the thesis. The findings from test phase one revealed that the implemented smile estimation algorithm could not differentiate between ambiguous smiles such as the one seen in Picture 19 among others. Therefore test phase two will comprise of pictures that clearly state a definitive smile or no smile. The extremes have to be found in order to prove if the algorithm can differentiate between smiles that are clear. This will unfortunately void the problem statement in the sense that the computer cannot be able to rate and interpret the level of smile with the present implemented algorithm on the same level and understanding as humans.
Table - Test Phase Two - Test Subjects
8.3.2.Test Results – ComputerThe same set of pictures as the test group where given to the computer for analysis. Table 4 shows the results and average for each picture. Each picture was again analysed three times and from that the average was calculated. As with test phase one, the average rating for each picture was divided by the region of interest (the area the algorithm uses for facial feature detection). The resulting calculation is the smile estimation (The spread sheet from test phase two, computer results, can be found in Appendix 16.4) Table - Test Phase Two - Computer Results
Table - Test Phase Two - Test Subjects, average rating - Graph The following graphs show the plotted data from test phase two – test subjects and test phase two – computer. The value of the x-axis corresponds to the picture number and the value of the y-axis corresponds to the rating. As can be seen on the graphs, by selecting pictures that perfectly match the abilities of the algorithm, the results by the computer and test subjects are almost equal. The algorithm therefore functions accordingly to being able to precisely estimate the level of smile in each picture. Although the estimations by the computer have not been normalised, the valleys and peaks from both graphs (discounting the intensity) follow the same path. Table - Test Phase Two - Computer Results - Graph Therefore compared to the results from test phase one, the results in test phase two show that the program was able to correctly estimate the level of smile in each picture as the graphs fluctuate the same. Directory: projekter -> files files -> IT’d be weird without mcdonald’S files -> Of Maj 2014 aau introduction & Theory files -> Representations of Gender in Agatha Christie’s Poirot and Miss Marple files -> Research in contemporary social movements: a case study of Guatemala 2015 files -> Anne Hviid-Pilgaard Master Thesis 31/05 2012 Table of Contents files -> Jakob Nors-Ganer Aalborg University 28-05-2015 files -> Abstract: Purpose files -> Security and trust in IoT/M2m cloud based platform files -> The impact of french and british colonial rule on the he system in cameroon from the perspective of students files -> Rumble in the jungle the ‘Blessing’ and ‘Curse’ of Mineral Wealth in the Congo Download 0.97 Mb. Share with your friends:
|
