Face pixel detection using evidential calibration and fusion

Due to legal reasons, faces on a given image may have to be blurred. This may be achieved by combining several information sources, which may provide information at different levels of granularity; for instance face detectors return bounding boxes corresponding to assumed positions of faces, whereas skin detectors may return pixel level information. A general, well-founded and efficient approach to combining box-based information sources was recently proposed in the context of pedestrian detection. This approach relies on evidence theory to calibrate and combine sources. In this paper, we apply this approach to combine face (rather than pedestrian) detectors, in order to obtain a state-of-the-art face blurring system based on multiple detectors. Then, we propose another approach to tackle the blurring problem, which consists essentially in applying at the pixel-level the central idea - combining evidentially calibrated information sources - of the preceding box-based approach. This shift of focus induces several conceptual advantages. In addition, the proposed approach shows better performances on a classical face dataset, as well as on a more challenging one.