Robust face detection by simple means

Abstract

Face detection is still one of the core problems in computer vision, especially in unconstrained real-world situations where variations in face pose or bad imaging conditions have to be handled. These problems are covered by recent benchmarks such as Face Detection Dataset and Benchmark (FDDB) [2], which reveals that established methods, e.g, Viola and Jones [8] suffer a drop in performance. More effective approaches exist, but are closed source and not publicly available. Thus, we propose a simple but effective detector that is available to the public. It combines Histograms of Orientated Gradient (HOG) [1] features with linear Support Vector Machine (SVM) classification. 2 Technical Details One important aspect in the training of our face detector is bootstrapping. Thus, we rely on iterative training. In particular, each iteration consists of first describing the face patches by HOGs [1] and then learning a linear SVM. At the end of each iteration we bootstrap with the preliminary detector hard examples to enrich the training set. We perform several bootstrapping rounds to improve the detector until the desired false positive per window rate is reached. Interestingly, we found out that picking up false positives at multiple scales in a sliding window fashion yields better results than just at a single scale. Testing several patch sizes and HOG layouts revealed that a patch size of 36 by 36 delivers the best results. For the HOG descriptor we ended up with a block size of 12x12, 4x4 for the cells. Prior to the actual training we gathered face crops of the Annotated facial landmarks in the wild (AFLW) dataset [4]. As AFLW includes the coarse face pose we are able to retrieve about 28k frontal faces by limiting the yaw angle between ± π 6 and mirroring them. For each face we crop a square region between forehead and chin. The non-face patches are obtained by randomly sampling at multiple scales of the PASCAL VOC 2007 dataset, excluding the persons subset. 3