Comparison of different face detection algoritms (in python) tested on realworld footage, but a the face was printed on paper.
Video files are available on my drive.
Detectors used: Haar cascade, Hog detector (from dlib library), DNN (from OpenCV library).
The detectors were tested with 3 different lightning conditions: Natural daily light, Dimmed daily light in the evening and Artificial industrial light.
The tests were conducted under 5 different camera angles (90°, 75°, 60°, 45°, 30°), where 90° means straight on.
For the evaluation process the F1 score was used, which is defined as following: F1 = 2*TP / (2*TP+FP+FN)
Detailed info can be found in the report.