Source-linked AI summary
Face Detection using Deep Learning: An Improved Faster RCNN Approach
Xudong Sun, Pengcheng Wu, Steven C. H. Hoi
TL;DR
Face detection is an important computer-vision problem, and the paper addresses it by extending Faster R-CNN with feature concatenation, hard-negative mining, pretraining, and multi-scale training. The resulting method achieved state-of-the-art performance on FDDB and ranked best among published approaches, including on ROC-curve evaluation.
Problem
Traditional face-detection pipelines depend on handcrafted features and separately optimized components, while face detection remains an important task for later face-related applications.
Method
The paper extends Faster R-CNN with feature concatenation, hard-negative mining, multi-scale training, model pretraining, and calibrated design parameters.
Results
The proposed method achieved state-of-the-art FDDB performance, ranking best among published approaches and scoring highest on the standard ROC curves.
Takeaways & Limitations
The combined strategies produced the best detection performance in the reported ablation experiments and enabled detection of difficult cases including occlusion, low resolution, extreme poses, and illumination.
Abstract
from arXiv · showhide
In this report, we present a new face detection scheme using deep learning and achieve the state-of-the-art detection performance on the well-known FDDB face detetion benchmark evaluation. In particular, we improve the state-of-the-art faster RCNN framework by combining a number of strategies, including feature concatenation, hard negative mining, multi-scale training, model pretraining, and proper calibration of key parameters. As a consequence, the proposed scheme obtained the state-of-the-art face detection performance, making it the best model in terms of ROC curves among all the published methods on the FDDB benchmark.
1 Introduction
Face detection is an important computer-vision problem supporting later face-related applications, but traditional pipelines rely on separately optimized handcrafted features and classifiers. The paper extends Faster R-CNN with several strategies and reports the best published performance on FDDB.
- Face detection supports applications including face verification, face recognition, and face clustering.
- Traditional approaches require expert-designed features and separately optimize pipeline components, often producing sub-optimal detection systems.
- Deep learning methods avoid handcrafted feature design and have achieved strong results across computer-vision tasks.
- Face detection can be treated as a specialized object-detection task, motivating extensions of RCNN and its variants.
- The proposed method extends Faster R-CNN using feature concatenation, hard negative mining, and multi-scale training.
- The method achieved state-of-the-art performance on FDDB, ranking best among published approaches.
2 Related Work
Earlier face detectors developed handcrafted features, classifiers, and multi-detector or part-based strategies, but robustness gains were limited or computationally expensive. Deep learning approaches subsequently improved face detection and increasingly adopted Faster R-CNN.
- The Viola–Jones framework used rectangular Haar-like features with a cascaded AdaBoost classifier for real-time face detection.
- Researchers introduced HOG, SIFT, SURF, ACF, NPD, SVM, and random-forest approaches to address feature complexity and detector design.
- Multiple detectors and deformable part models were used to improve robustness across views, poses, and expressions.
- These traditional robustness strategies were usually more time-consuming, while their detection-performance improvements were relatively limited.
- Deep learning methods often significantly outperform traditional computer-vision methods, and Faster R-CNN achieved promising face-detection results.
3 Our Approach
The approach extends Faster R-CNN with a staged training procedure and architectural changes for face detection. It combines pretraining, hard-negative mining, feature concatenation, and multi-scale training to improve handling of difficult and varied face appearances.
- Overview of Methodology: Faster R-CNN combines an RPN for generating likely object regions with Fast R-CNN for classification and boundary refinement.
- Overview of Methodology: The training procedure pretrains on WIDER FACE, harvests hard negatives, fine-tunes on FDDB, and applies multi-scale training and feature concatenation.
- Feature Concatenation: Feature concatenation pools feature maps from multiple convolutional layers to capture both fine-grained and high-level RoI information.
- Hard Negative Mining: Hard negatives are regions where the network made incorrect predictions; feeding them back reinforces training toward fewer false positives and better classification.
- Hard Negative Mining: Hard negatives are selected using an IoU threshold below 0.5 and explicitly added to RoIs during fine-tuning.
- Multi-Scale Training: Multi-scale training randomly assigns one of three image scales, helping the detector learn across varied face sizes and become more robust to scale.
4 Experiments
Experiments evaluated the proposed face detector on FDDB using cross-validation, qualitative examples, and ablation studies. The results showed state-of-the-art benchmark performance and identified contributions from anchor-size changes, pretraining with hard-negative mining, feature concatenation, and multi-scale training.
- Experimental Setup: The FDDB benchmark contains 5,171 faces in 2,845 images with occlusions, difficult poses, low resolution, and out-of-focus faces.
- Experimental Setup: The evaluation used 10-fold cross-validation with horizontal flipping and random resizing to shorter-side scales of 480, 600, or 750.
- FDDB Benchmark Results: The submitted model achieved the highest standard ROC scores among published FDDB methods, clearly outperforming the second-highest method on continuous ROC.
- FDDB Benchmark Results: Qualitative results included detections of non-frontal, heavily occluded, low-resolution, and extreme-pose or illumination faces; many selected false positives appeared to be missing annotations.
- Ablation Experiments: Adding a 64 × 64 anchor group increased RPN anchors from 9 to 12 and allowed detection of more small boxes.
- Ablation Experiments: WIDER FACE pretraining improved recall but increased false positives, while hard-negative mining significantly reduced those false positives.
- Ablation Experiments: Feature concatenation improved classification by combining multiple-layer features, and random-scaling multi-scale training further increased detection performance.
- Ablation Experiments: Combining all evaluated strategies produced the best detection performance in experiment ID 7.
5 Conclusions
The paper extends Faster RCNN with several strategies for face detection and reports state-of-the-art results on the FDDB benchmark.
- The method extends the state-of-the-art Faster RCNN framework for face detection using deep learning techniques.
- The proposed improvements include feature concatenation, multi-scale training, hard negative mining, and proper configuration of RPN anchor sizes.
- The method achieved state-of-the-art results on the FDDB face detection benchmark, ranking best among published methods.
- Future work will address the method's efficiency and scalability for real-time face detection.