BookClub artistic makeup and occlusions face data

The novel BookClub artistic makeup data set contains images of 21 subjects. Each subject’s data may contain a photo-session series of photos with no-makeup, various makeup, and images with other obstacles for facial recognition, such as wigs, glasses, jewellery, face masks, or various types of headdress. Overall, the data set features 37 photo-sessions without makeup or occlusions, 40 makeup sessions, and 17 sessions with occlusions. Each photo-session contains circa 168 RAW images of up to 4288×2848 dimensions (available by request) of six basic emotional expressions (sadness, happiness, surprise, fear, anger, disgust), a neutral expression, and the closed eyes photo-shoots taken with seven head rotations at three exposure times on the off-white background. Default publicly available downloadable format is a JPEG of the 1072×712 resolution. The subjects’ age varies from their twenties to sixties. Race of the subjects is predominately Caucasian and some Asian. Gender is approximately evenly split between sessions. The photos were taken over the course of two months. A few sessions were done later, and some subjects posed at multiple sessions over several week intervals in various clothing with changed hairstyles. The motivation for the novel BookClub data set creation was an attempt to fill in the apparent gap in the makeup and occlusions data sets of high quality and resolution images featuring sophisticated original artistic makeup and non-trivial occlusions, made in the controlled environment with a wide variety of the represented emotions, orientations, lighting conditions per subject, for subjects representing age ranges, genders, and race. To emulate the real-life conditions, when makeup or other occlusions of the subjects may not be predicted and included into training sets beforehand, non-makeup and non-occlusion sessions were selected into the training set, which amounted to roughly 6200 images. Then, the facial recognition accuracy estimation was run with the same as above mentioned training parameters of AlexNet for each makeup and occlusion photo-session. While the majority of the sessions, were correctly identified, scoring high 93−100% accuracy, some sessions scored significantly lower accuracy in the 50−90% range, and some sessions were misidentified. Some sessions with especially high contrast, dark pigment makeup, or created by professional artists scored close to 0% accuracy. Even worse, some of the low accuracy makeup scored high guess scores for the wrong subject, it appears making accuracy < 90% effectively unreliable by upper bound. Experiments on the novel BookClub data set of artistic makeup and occlusions have demonstrated that even contemporary state-of-the-art face recognition algorithms such as AlexNet (even retrained on a large amount of data) when meeting with real-life obstacles, perform poorly at the desired high confidence levels. v.3 update: Subjects' age and emotion partition (bc2e) are added.