WIDER FACE: A Face Detection Benchmark (1511.06523v1)

Abstract: Face detection is one of the most studied topics in the computer vision community. Much of the progresses have been made by the availability of face detection benchmark datasets. We show that there is a gap between current face detection performance and the real world requirements. To facilitate future face detection research, we introduce the WIDER FACE dataset, which is 10 times larger than existing datasets. The dataset contains rich annotations, including occlusions, poses, event categories, and face bounding boxes. Faces in the proposed dataset are extremely challenging due to large variations in scale, pose and occlusion, as shown in Fig. 1. Furthermore, we show that WIDER FACE dataset is an effective training source for face detection. We benchmark several representative detection systems, providing an overview of state-of-the-art performance and propose a solution to deal with large scale variation. Finally, we discuss common failure cases that worth to be further investigated. Dataset can be downloaded at: mmlab.ie.cuhk.edu.hk/projects/WIDERFace

• The paper introduces a large-scale dataset with over 393,000 annotated faces to challenge and improve current face detection methods.
• It provides a rigorous benchmark by analyzing state-of-the-art detectors across varying conditions like scale, occlusion, and pose.
• The work proposes a multi-scale two-stage cascade framework that significantly enhances detection performance on challenging face scenarios.

WIDER FACE: A Face Detection Benchmark

The paper "WIDER FACE: A Face Detection Benchmark" addresses the progress and existing challenges in the domain of face detection within the computer vision community. Presented by Shuo Yang, Ping Luo, Chen Change Loy, and Xiaoou Tang, the paper introduces a novel and extensive dataset named WIDER FACE, which is significantly larger and more varied than existing face detection datasets.

Contributions of WIDER FACE

The primary contributions of the paper are manifold:

1. Introduction of a Large-scale Dataset: The WIDER FACE dataset comprises 32,203 images with 393,703 labeled faces, making it ten times larger than the largest pre-existing face detection dataset. The dataset includes a wide range of facial variations in terms of pose, scale, occlusion, expression, appearance, and illumination. This level of variability is intended to bridge the gap between current face detection performance and real-world requirements.
2. Rich Annotations: The dataset is meticulously annotated with attributes such as occlusion, pose, event categories, and face bounding boxes. This rich annotation allows for in-depth analysis and helps identify specific failure modes of existing algorithms.
3. Benchmarking and Analysis: The authors benchmark several state-of-the-art face detection systems using the WIDER FACE dataset. They analyze their performance under different conditions such as variations in scale, occlusion levels, and pose deformations. This comprehensive benchmarking provides insights into the strengths and weaknesses of current algorithms.
4. Multi-scale Two-stage Cascade Framework: The paper proposes a novel detection framework to handle large scale variations effectively. This framework involves training multiple convolutional networks where each network is specialized in detecting faces of a specific scale range. By dividing the face detection task into manageable sub-tasks, the proposed framework aims to enhance detection performance, particularly for faces with significant scale variations.

Experimental Results

The experimental section provides a thorough evaluation of existing face detection methods on the WIDER FACE dataset. Key findings include:

• Performance Analysis: Modern detectors like Faceness, DPM, and ACF are benchmarked on easy, medium, and hard subsets of the dataset. The experiments reveal that while these methods perform reasonably well on large and unoccluded faces, their performance degrades significantly on small, occluded, or atypically posed faces.
• Retraining with WIDER FACE: The paper demonstrates that retraining existing detectors on the WIDER FACE dataset can substantially improve their performance. For instance, retrained versions of the ACF and Faceness detectors show marked improvements on both WIDER FACE and FDDB datasets.
• Proposal of Multi-scale Cascade: The proposed multi-scale cascade CNN outperforms baseline methods, particularly on the hard subset of the WIDER FACE dataset. This result underscores the importance of handling multi-scale variations effectively in face detection tasks.

Implications and Future Directions

The introduction of the WIDER FACE dataset has significant practical and theoretical implications:

• Practical Implications: For applications such as surveillance and security, where faces are often captured in challenging conditions (e.g., occlusions, small scales, extreme poses), the WIDER FACE dataset provides a more realistic benchmark. This can drive the development of more robust face detection systems tailored for real-world scenarios.
• Theoretical Implications: From a research perspective, the dataset facilitates a deeper understanding of the limitations of current approaches. It prompts the exploration of novel ideas and methodologies for tackling the persistent challenges in face detection.

Conclusion

In conclusion, the WIDER FACE dataset represents a substantial advancement in the resources available for face detection research. By offering a much larger and more varied set of labeled images, it sets a new standard for benchmarking face detection algorithms. The insights derived from the dataset underscore the need for continued innovation in handling small scale, occluded, and atypical faces. Future research in AI and computer vision can leverage this dataset to develop more accurate and generalizable face detection systems, ultimately pushing the boundaries of what is possible in the field.