Week 12 Journal

This week, Jennifer and I focused on collected data. First, we tested the accuracy between different amounts of face detection (using face tracking whenever face detection was not available). Face tracking works faster than face detection because it sorts through keypoints instead of passing all the Haar cascade filters to find a face. We conducted 10 trials for 3 cases - for a face being detected about 6 times, 3 times, and 1 time. From this data, we found that having a face detection about 2-3 times during the data collection period was beneficial - it allows for more data collection and makes sure that the face does not move too much during the data collection period.

One thing we noticed while using face tracking was that the keypoints that were being detected were very sensitive to the background. From this, we learned that the background had to be plain, and could not have any distracting elements (like a doorknob). Repositioning the camera helped acquire more data.

Finally, we realized that our code depended on finding the mean RGB values from converting the mean LAB values. This conversion may not be accurate. Instead of depending on this conversion, we found the mean RGB values of the region of interest by parsing through each pixel's RGB values and finding the mean of each channel. Although this added to computational complexity of our algorithm, it allowed for more distinct changes in the RGB channels. Previously, the conversion would allow only integers, but manually calculating the mean allowed for floats.

We spent the rest of the week cleaning up our code, writing our final report, and working on our presentation.