TESTING
July 12th, 2019
For the initial set of results, testing was done on two trained subjects, Iqbal and Catalina. Using a training set of between 20-30 frontal face pictures, the software was trained for each subject. In addition, our third subject, Iqbal’s roommate, Charlie was used as a stranger to the system.
For Iqbal, 20 trails were conducted under optimal lighting with glasses and without glasses. Another 20 trails were conducted with no lighting to simulate a nighttime environment. Trails for the last three conditions, optimal lighting wearing a hat, adequate lighting with water, and wearing sunglasses, were conducted 10 times each. These first three conditions are more often the case, so they were conducted 20 times as opposed to the 10 times for the latter 3 conditions.
July 13th, 2019
For the next subject, Catalina, the same conditions were tested except for the condition ‘Optimal Lighting wearing a Hat.’ This was removed because of time constraints and because it was clear from the first set of tests that wearing a hat severely diminished the program’s ability to both detect and recognize a face. The same number of trials from the first subject were conducted and the outcomes for each trial and corresponding condition are shown below the video.
July 14th, 2019
The complete system had to be tested with a third subject, a stranger. This stranger would be Iqbal’s roommate, Charlie. Unfortunately, during this test there were many false positives and the door would actually open for the stranger. After the first round of tests were done on Charlie, we made some adjustments to the confidence or distance required for the LBPH recognizer. It was changed so that the final distance or variable, ‘conf’, was set to below 145 to accept a face. This helped eliminate the false positives significantly. The table with Charlie’s success and failure outcomes are displayed below the video.