PV080 Vladimír Bouček Spring Semester 2021 UČO: 492927 Assignment 4 – face detection 1. Hypothesis – gaussian noise Hypothesis: Applying 50% gaussian noise lowers face detection confidence rate at least by 50 %. In first hypotheses I tried to test the effect of gaussian noise on face detection. I started with baseline photo and then I continuously added 5 % to next photos. I used Adobe Photoshop for editing images in first three hypothesis. In table below we can see that face detection certainty gets continuously lower with every 5 % added. At 35 % it rapidly drops, and it starts to detect another face in room. At 45 % (and more) I was unable to detect face at all, however with human eye you can clearly spot a face. This result support my hypothesis and thus I cannot reject it. Gaussian noise intensity Face detection confidence Baseline photo 0.9999995 5 % 0.9999993 10 % 0.9999976 15 % 0.9999918 20 % 0.99985635 25 % 0.99466693 30 % 0.9712634 35 % 0.60844094, 0.2375087 40 % 0.27032733,0.24382938 45 % No face detected 50 % No face detected 55 % No face detected 60 % No face detected Face detection of image before and after adding 40 % gaussian noise PV080 Vladimír Bouček Spring Semester 2021 UČO: 492927 2. Hypothesis – gaussian blur Hypothesis: Adding gaussian blur with radius 30 will lower confidence level at least by 50 %. In second hypothesis I tried an “opposite” to first hypothesis, instead of adding noise, I tried to add gaussian blur. I used same baseline photo and then I continually increased blur up to radius 30. I thought that face detection would not be vulnerable to gaussian blur as it was to noise, but even with radius 30 it was still able to find my face with more than 99% confidence. It basically didn’t affect detection at all. Based on this result, I have to reject my hypothesis. Gaussian blur radius Face detection confidence Baseline photo 0.9999995 3 0.9999995 6 0.9999994 9 0.99999917 12 0.9999989 15 0.9999982 18 0.9999968 21 0.99999475 24 0.9999918 27 0.9999869 30 0.9999783 Face detection of image before and after adding gaussian blur with radius 30 3. Hypothesis – exposure Hypothesis: Increasing exposure by 3 or lowering by 6 will lower face detection confidence level at least by 25 %. In third hypothesis I tried focus on exposure adjusting and its effect on face detection. As we can see in the table below, exposure has minimal impact. I had to make the photo almost completely black/white for face detection to fail. But I have to reject my hypothesis, because mentioned values didn’t effect detection at all. PV080 Vladimír Bouček Spring Semester 2021 UČO: 492927 Exposure Face detection rate -7.0 0.49925607 -6.0 0.99513006 -5.0 0.9999281 -4.0 0.9999924 -3.0 0.9999976 -2.0 0.9999988 -1.0 0.9999993 baseline 0.9999995 +1.0 0.9999993 +2.0 0.99999845 +3.0 0.99994826, 0.29353362 +4.0 0.9983045,0.3189708 Face detection before and after increasing exposure by 3 4. Hypothesis – respirator Hypothesis: Face detection software won’t find my face while wearing a respirator at least in 25 % of test photos. Because the covid pandemic is still not over, I got an idea to try wearing a respirator. I took 10 photos without respirator as a baseline and then 10 photos with respirator. I tried do this from same angles and distance, but it is not 100% accurate, because I don’t have a tripod. Result is quite surprising to me - as we can see in table below, it was still able to find my face on all photos. On some photos it detected another face in the room, but this happened on photos without respirator as well and it was still able to find my face, so I count it as a success in this hypothesis. Since it successfully found my face on all photos, I have to reject my hypothesis. PV080 Vladimír Bouček Spring Semester 2021 UČO: 492927 Photo Face detection rate Photo with respirator Face detection rate close 0.99999416 resp_close 0.9915976,0.39403212 close_left 0.99988055 resp_close_left 0.9544062,0.27729174 close_right 0.9991843,0.32237703 resp_close_right 0.87236357 close_up 0.999866 resp_close_up 0.931553 close_down 0.99848276 resp_close_down 0.98325515,0.20802407 far 0.9953412 resp_far 0.99770665 far_left 0.9478564 resp_far_left 0.9490679 far_right 0.96785206,0.318539 resp_far_right 0.89517605 far_up 0.9942842 resp_far_up 0.5028247 far_down 0.9999516 resp_far_down 0.94568545,0.38675725 Face detection on images far.jpeg and resp_far.jpeg 5. Hypothesis – sunglasses Hypothesis: Face detection software won’t find my face while wearing sunglasses in at least 25 % of test photos. Because covering my mouth and nose didn’t fool the program, I got an idea to use sunglasses to cover my eyes instead. Thus, I stated a similar hypothesis, but this time I put on a pair of sunglasses. I used same baseline photoset as in the fourth hypothesis. The results are quite surprising to me, because it was still able to find my face on all photos, thus I have to reject my hypothesis. Photo Face detection rate Photo with sunglasses Face detection rate close 0.99999416 glasses_close 0.99999166 close_left 0.99988055 glasses_close_left 0.9999763 close_right 0.9991843,0.32237703 glasses_close_right 0.9990709 close_up 0.999866 glasses_close_up 0.40867317 close_down 0.99848276 glasses_close_down 0.9996362 far 0.9953412 glasses_far 0.9991321 far_left 0.9478564 glasses_far_left 94978935,0.20662367 PV080 Vladimír Bouček Spring Semester 2021 UČO: 492927 far_right 0.96785206,0.318539 glasses_far_right 0.9222849 far_up 0.9942842 glasses_far_up 0.2974371 far_down 0.9999516 glasses_far_down 0.9976871 Face detection on close_left.jpeg and glasses_close_left.jpeg Summary From my observation I think that face detection is mostly vulnerable to randomness, because it had biggest problem with noise. I think that this could be achieved with some random very unusual make up for example, but I don’t have resources to test this. It didn’t have almost any problem with exposure and other changes, that effects whole image (unless it got to the extreme like almost black image). It also didn’t have any problem with partial face covering, it was still able to detect face with high confidence level. I think that AI behind this program could have been trained with some similar photos with glasses/masks as well. However, face detection software had problems too. It was often detecting “face” in room or in my glasses, I don’t really understand why this is happening. Only explanation that came to my mind is that threshold for accepting part of an image as a face is set way too low.