"Unconstrained face detection is the task of robustly finding and locating faces in an image subject to possible variations in facial scale, blur, pose, illumination, occlusion, and facial expression. It is a critical first step towards a host of modern surveillance applications, including but not limited to face verification, face recognition, face tracking, and human-computer interaction. Though much progress has been made in unconstrained face detection during the past decade, the majority of work focuses on improving the detection robustness on variations caused by blur, pose, illumination, occlusion and facial expression. Facial scale, despite its immense influence on face detection accuracy, has received much less attention than have the above factors. This is partially due to the fact that most traditional face detection benchmark datasets tend to collect faces of relatively large size and with modest scale variation. Nonetheless, in real-world applications, such as surveillance systems, it is imperative to possess an equal ability to detect both big faces (close to camera) and tiny ones (far away from the camera) at the same time. To the best of our knowledge, no published face detection algorithm can detect a face as large as 1000 x 1000 pixels while simultaneously detecting another one as small as 10 x 10 pixels within a single image with similarly high accuracy.We introduce a Multi-Path Face Detection Network (MP-FDN) to filter an image for simultaneously proposing and verifying different sized faces in parallel paths. This is the first time that faces across a large span of scales are detected by a single network with forked detection paths. More importantly, the division of the paths are not handcrafted, but totally based on the scale sensitivity inherent in the convolutional networks that was also discovered in this thesis for the first time. MP-FDN consists of two stages. The first stage is a Multi-Path Face Proposal Network (MP-FPN) that suggests faces at three different scale ranges. This design is based on our observation that the hierarchical multi-scale layers of deep convolutional networks (ConvNet) can inherently represent face patterns at multiple scales. In particular, low-level ConvNet layers are more sensitive to tiny faces, while high-level ConvNet layers are more discriminative to big faces. To this end, MP-FPN utilizes three parallel outputs of the convolutional feature maps to simultaneously predict small, medium and large candidate face regions, respectively. The second stage is a Multi-Path Face Verification Network (MP-FVN) that further eliminates false positives while including false negatives. MP-FVN utilizes the same three parallel paths as MP-FPN. For each detection path, it pools features from both a face candidate region (provided by MP-FPN) and a larger contextual region (surrounding the face candidate region). These facial and contextual features are then concatenated to provide a more accurate "faceness" probability to the face candidate. Note that the network structure and hyper-parameters of MP-FPN and MP-FVN are completely based on controlled experiments, rather than being "handcrafted". To testify to the performance of MP-FDN on the basis its ability to perform face detection, we conducted comprehensive experiments on two challenging public face detection benchmark datasets: WIDER FACE and FDDB datasets. MP-FDN consistently achieves better than the state-of-the-art performance on both of them. Specifically, on the most challenging so-called "hard partition" of WIDER FACE test set that contains faces as small as about 9 pixels and as large as more than 1000 pixels in height, MP-FDN outperforms the former best result by 9.8% for the Average Precision. This demonstrates that MP-FDN is a viable and accurate face detector for unconstrained face detection, especially in the case of large scale variations." --