Source-linked AI summary
FaceBoxes: A CPU Real-time Face Detector with High Accuracy
Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, Stan Z. Li
TL;DR
FaceBoxes targets the unresolved challenge of combining high face-detection accuracy with real-time CPU speed. It uses a lightweight single-network architecture with RDCL, MSCL, and anchor densification, and reports real-time VGA performance with state-of-the-art benchmark results.
Problem
Face detection still needs to achieve real-time CPU speed while maintaining high performance because effective models are computationally prohibitive.
Method
FaceBoxes combines RDCL for CPU efficiency, MSCL for multi-scale face handling, and anchor densification for balanced anchor density and small-face recall.
Results
20 FPS on a single CPU core and 125 FPS on a GPU for VGA-resolution images, with state-of-the-art accuracy reported on AFW, PASCAL face, and FDDB.
Takeaways & Limitations
FaceBoxes provides a lightweight face detector whose speed is invariant to the number of faces in an image while addressing multiple face scales.
Abstract
from arXiv · showhide
Although tremendous strides have been made in face detection, one of the remaining open challenges is to achieve real-time speed on the CPU as well as maintain high performance, since effective models for face detection tend to be computationally prohibitive. To address this challenge, we propose a novel face detector, named FaceBoxes, with superior performance on both speed and accuracy. Specifically, our method has a lightweight yet powerful network structure that consists of the Rapidly Digested Convolutional Layers (RDCL) and the Multiple Scale Convolutional Layers (MSCL). The RDCL is designed to enable FaceBoxes to achieve real-time speed on the CPU. The MSCL aims at enriching the receptive fields and discretizing anchors over different layers to handle faces of various scales. Besides, we propose a new anchor densification strategy to make different types of anchors have the same density on the image, which significantly improves the recall rate of small faces. As a consequence, the proposed detector runs at 20 FPS on a single CPU core and 125 FPS using a GPU for VGA-resolution images. Moreover, the speed of FaceBoxes is invariant to the number of faces. We comprehensively evaluate this method and present state-of-the-art detection performance on several face detection benchmark datasets, including the AFW, PASCAL face, and FDDB. Code is available at https://github.com/sfzhang15/FaceBoxes
1. Introduction
FaceBoxes addresses the CPU face-detection trade-off between accuracy and speed with a lightweight single-network design, targeting real-time detection while handling varied face scales. Its contributions include RDCL, MSCL, anchor densification, and benchmark evaluation.
- Motivation: CPU face detection remains difficult because accurate classification amid visual variation and searching many positions and sizes impose conflicting accuracy and efficiency requirements.High-accuracy detectors tend to be computationally expensive, making real-time CPU performance an open practical issue.
- Motivation: Cascaded CNN detectors improve performance and efficiency but slow as face count increases, complicate training, and achieve about 14 FPS on VGA images.These limitations motivate a single end-to-end network with runtime independent of the number of faces.
- FaceBoxes: FaceBoxes uses one fully convolutional network containing RDCL and MSCL, enabling end-to-end training while targeting CPU real-time speed and multiple face scales.RDCL supports speed; MSCL enriches receptive fields and distributes anchors across layers with different resolutions.
- FaceBoxes: The paper proposes anchor densification so different anchor types share image density, improving recall for small faces.This strategy is presented as a core contribution alongside the network architecture.
- Evaluation: The paper reports improved state-of-the-art performance on the AFW, PASCAL face, and FDDB datasets.These datasets are the stated evaluation benchmarks for the proposed detector.
2. Related work
Prior face detection methods developed along hand-crafted-feature and CNN-based directions. Hand-crafted approaches emphasized speed and engineered features, while CNN methods improved robustness and detection performance.
- Method categories: Face detection methods are broadly divided into hand-crafted-feature systems and CNN-based systems.The related-work section organizes prior research around these two categories.
- Hand-crafted features: Viola-Jones and subsequent hand-crafted methods combined engineered features, boosting or classifiers, and cascade structures for real-time detection.Later work also introduced local features, boosting algorithms, and new cascade structures.
- Hand-crafted features: Deformable part model methods used supervised parts, pose partitioning, improved training, or efficient inference to achieve strong detection performance.These methods extended the structural-model direction beyond cascade frameworks.
- CNN-based methods: CNN face detectors progressed from sliding-window and orientation-specific networks to feature extraction, facial-attribute modeling, and CNN-based cascades.The cited examples span early CNN face detection and more recent architectures.
3. FaceBoxes
FaceBoxes combines RDCL, MSCL, and anchor densification to make face detection accurate and efficient on CPU devices. Its design rapidly reduces computation, handles multiple face scales, balances anchor density, and specifies an end-to-end training pipeline.
- FaceBoxes: The method is presented as three contributions intended to make FaceBoxes accurate and efficient on CPU devices.These contributions are RDCL, MSCL, and anchor densification.
- Rapidly Digested Convolutional Layers: RDCL rapidly shrinks feature-map spatial size to enable real-time CPU face detection.Its convolution and pooling strides total 32, reducing input spatial size by 32 times.
- Multiple Scale Convolutional Layers: MSCL handles different face scales by combining multi-scale feature maps with receptive fields enriched by Inception modules.Anchors are associated with Inception3, Conv3 2, and Conv4 2, while different convolution branches capture varying receptive-field sizes.
- Anchor densification strategy: Anchor densification equalizes anchor density across scales by adding 4 placements for 32×32 anchors and 2 for 64×64 anchors.The strategy produces density 4 for different anchor scales, helping various face sizes match almost the same number of anchors.
- Training: FaceBoxes training includes augmentation, anchor matching, RPN-style classification and regression losses, hard-negative mining, and SGD optimization.Anchors are matched using best overlap and a 0.35 Jaccard threshold; negative-to-positive examples are capped at 3:1.
4. Experiments
FaceBoxes is evaluated for runtime efficiency, component contributions, and benchmark performance. The experiments report real-time CPU inference, gains from the proposed components, and strong results across AFW, PASCAL face, and FDDB.
- Runtime efficiency: 20 FPS on a CPU and 125 FPS on a GPU are reported for VGA-resolution images, with 4.1MB model size.The CPU and GPU measurements use the stated inference setup and compare FaceBoxes with recent CNN-based methods.
- Model analysis: Ablation experiments on FDDB remove or replace anchor densification, MSCL, and C.ReLU to assess their effects on accuracy and speed.The experiments use the same settings except for specified component changes.
- Model analysis: Removing anchor densification reduces FDDB mAP from 96.0% to 94.9%, a 1.1% decline attributed to reduced small-face recall.The strategy densifies 32 × 32 and 64 × 64 anchors to improve recall for small faces.
- Model analysis: MSCL increases FDDB mAP by 1.0%, while RDCL decreases mAP by 0.1% for about 19.3ms faster inference.The reported effects are associated with diverse receptive fields, multi-scale anchor tiling, and RDCL efficiency.
- Evaluation on benchmark: FaceBoxes outperforms competing methods by a large margin on AFW, PASCAL face, and FDDB evaluations.The reported comparisons include precision-recall curves on AFW and PASCAL face and discontinuous and continuous ROC curves on FDDB.
5. Conclusion
FaceBoxes addresses the challenge of combining CPU real-time speed with high face-detection performance through RDCL, MSCL, and anchor densification. The detector achieves state-of-the-art benchmark performance while reaching 20 FPS on CPU and 125 FPS on GPU for VGA-resolution images.
- Conclusion: FaceBoxes combines RDCL for CPU speed, MSCL for multi-scale faces, and anchor densification for improved small-face recall.The network is described as lightweight and powerful, with the proposed components contributing to accuracy and efficiency.
- Conclusion: 20 FPS on CPU and 125 FPS on GPU are reported for VGA-resolution images, alongside state-of-the-art performance on common face-detection benchmarks.The conclusion names AFW, PASCAL face, and FDDB as the benchmark datasets evaluated in the paper context.