The WIDER FACE dataset is a large-scale face detection benchmark containing 32,203 images with 393,703 labeled face bounding boxes. It is ten times larger than existing datasets and includes rich annotations for occlusion, pose, event categories, and face bounding boxes. The dataset is challenging due to variations in scale, pose, occlusion, and background clutter. It is designed to evaluate face detection algorithms under real-world conditions, including heavy occlusion, small scale, and atypical poses. The dataset is used to benchmark several face detection systems, showing that current algorithms struggle with these challenges. The WIDER FACE dataset is also an effective training source for face detection, as demonstrated by improved performance on benchmark tasks. The paper introduces a multi-scale two-stage cascade framework for face detection, which uses a divide-and-conquer strategy to handle large-scale variations. The framework includes a set of convolutional networks trained to detect faces in specific scale ranges. The paper also evaluates the performance of four representative face detection algorithms on the WIDER FACE dataset, showing that current methods have limited success in handling challenging scenarios. The results indicate that the WIDER FACE dataset is a valuable resource for advancing face detection research.The WIDER FACE dataset is a large-scale face detection benchmark containing 32,203 images with 393,703 labeled face bounding boxes. It is ten times larger than existing datasets and includes rich annotations for occlusion, pose, event categories, and face bounding boxes. The dataset is challenging due to variations in scale, pose, occlusion, and background clutter. It is designed to evaluate face detection algorithms under real-world conditions, including heavy occlusion, small scale, and atypical poses. The dataset is used to benchmark several face detection systems, showing that current algorithms struggle with these challenges. The WIDER FACE dataset is also an effective training source for face detection, as demonstrated by improved performance on benchmark tasks. The paper introduces a multi-scale two-stage cascade framework for face detection, which uses a divide-and-conquer strategy to handle large-scale variations. The framework includes a set of convolutional networks trained to detect faces in specific scale ranges. The paper also evaluates the performance of four representative face detection algorithms on the WIDER FACE dataset, showing that current methods have limited success in handling challenging scenarios. The results indicate that the WIDER FACE dataset is a valuable resource for advancing face detection research.