The authors of the WIDER FACE dataset delve into the extensively researched domain of face detection in the field of computer vision. They selected a set of 32,203 images and labeled a total of 393,703 faces. The dataset is structured according to 61 distinct event classes. Within each event class, we adopted a random split of 40% for train, 10% for val, and 50% for test. The authors mention that a discrepancy exists between the current performance of face detection systems and the actual demands of real-world scenarios. To bridge this gap and to propel future face detection research, the authors introduce the WIDER FACE dataset, a substantial expansion over existing datasets with a tenfold increase in scale. This dataset is rich in annotations, incorporating attributes such as occlusions, poses, event categories, and bounding boxes for faces.

The dataset’s content includes faces that pose extreme challenges due to considerable variations in scale, pose, and occlusion. Furthermore, the authors demonstrate the efficacy of the WIDER FACE dataset as a robust training source for advancing face detection research. They evaluate several representative face detection systems, thereby offering an overview of the current state-of-the-art performance. Additionally, the authors propose a solution to address the challenge of accommodating large-scale variations. Lastly, they explore common failure cases that warrant further investigation.

The process of face detection is pivotal for various facial analysis algorithms, encompassing tasks such as face alignment, face recognition, face verification, and face parsing. The core objective of face detection is to discern the presence of faces within an image, and if detected, to localize the position and extent of each face. This seemingly straightforward task for humans proves to be immensely challenging for computational systems. The complexity associated with face detection stems from variations in factors such as pose, scale, facial expression, occlusion, blur, and illumination.

The WIDER FACE dataset, currently the most extensive face detection dataset, is chosen from the publicly accessible WIDER dataset. This dataset selection encompasses 32,203 images and encompasses 393,703 labeled faces. Organized into 60 event classes, the dataset’s images are divided into train, val, and test subsets for each event class. Two specific training/testing scenarios are outlined: Scenario-Ext, where an external face detector is trained and tested on the WIDER FACE test partition, and Scenario-Int, where a face detector is trained using the WIDER FACE training/validation partitions and subsequently tested on the WIDER FACE test partition. The evaluation metric follows that of the PASCAL VOC dataset.

The collection methodology for the WIDER FACE dataset is a subset of the larger WIDER dataset. Images were acquired through a multi-step process, involving the definition of event categories based on the Large Scale Ontology for Multimedia (LSCOM), retrieval of images using search engines, and manual cleaning of the dataset. The data annotation process involves bounding box labeling for all recognizable faces within the dataset. The bounding boxes are required to encompass the forehead, chin, and cheeks. Even when a face is occluded, it is labeled with a bounding box, along with an estimate of the occlusion scale. Faces deemed challenging for recognition due to low resolution and small scale receive an invalid flag. After annotating bounding boxes, attributes such as pose and occlusion level are further annotated. These annotations are conducted by a single annotator and cross-validated by two different individuals.

The distinctive properties of the WIDER FACE dataset present notable challenges due to the extensive variations in scale, occlusion, pose, and background complexity. These characteristics mirror the real-world requirements, providing a benchmark that’s more reflective of practical scenarios. The authors adopt generic object proposal techniques to assess the dataset’s difficulty and potential detection performance. This assessment reveals a lower detection rate for WIDER FACE compared to other face detection datasets, indicating its higher level of challenge. The dataset is classified into three difficulty levels: “Easy”, “Medium”, and “Hard”, based on the detection rate.