KTH Matematik 
|
| KTH Matematik |
|
 |
|
Tid: 25 maj 2018 kl 10.25-11.15. Seminarierummet F11, Institutionen för matematik, KTH, Lindstedtsvägen 22. Karta!Föredragshållare: Mattias Herlitz Title: Analyzing the Tobii Real-world-mapping tool and improving its workflow using Random Forests Abstract: The Tobii Pro Glasses 2 are used to record gaze data that is used for market research or scientific experiments. To make extraction of relevant statistics more efficient, the gaze points in the recorded video are mapped to a static snapshot with areas of interests (AOIs). The most important statistics revolve around fixations. A fixation is when a person is keeping his or her vision still for a short period of time. The method most used today is to manually map the gaze points. However, a faster method is automated mapping using the Real World Mapping (RWM) tool. In order to examine the reliability of RWM, the fixations from different recordings and projects were analyzed using Decision Trees. Further, a Random Forest (RF) model was constructed in order to predict if a gaze point was correctly or incorrectly mapped. It was shown that fixation classification on data from RWM performed significantly worse than when the same fixation classification on manually mapped data was run. It was shown that RWM works better when head movement is low and AOIs are set appropriately. This can guide researchers in setting up experiments, although major improvements of RWM is needed. The RF classifier showed promising results on several test sets for mapped gaze points. It also showed promising results for gaze points that were not mapped and were close in time to being mapped. In conclusion, the RF should replace current methods of estimating the quality of RWM gaze points. Gaze points that are classified as badly mapped can be manually remapped. If RWM fails to map large segments of gaze points to a snapshot, visually classifying these to be remapped is the preferred method. |
 |
| Sidansvarig: Jimmy Olsson Uppdaterad: 18/5-2018 |