Source-linked AI summary
Face Detection with the Faster R-CNN
Huaizu Jiang, Erik Learned-Miller
TL;DR
Face detection requires effective evaluation across established and newer benchmarks, while generic region proposals can impose substantial computational costs. This paper trains Faster R-CNN on WIDER, using a shared RPN and Fast R-CNN pipeline, and reports state-of-the-art performance on FDDB and IJB-A. The authors attribute its effectiveness to the RPN and shared convolutional layers, while noting that face-specific patterns may further improve results.
Problem
Generic object proposals rely on hand-crafted methods whose generation can dominate processing time, motivating learned, efficient proposal generation for face detection.
Method
The paper trains Faster R-CNN on WIDER, using an RPN and Fast R-CNN detector that share convolutional layers, with anchors spanning three scales and three aspect ratios.
Results
Faster R-CNN achieves state-of-the-art face detection performance on FDDB and IJB-A, outperforming all compared methods on IJB-A by a large margin.
Takeaways & Limitations
The reported effectiveness is attributed to the RPN and shared convolutional layers, which enable deep CNN proposals without extra computational burden.
Takeaways & Limitations
Although generic Faster R-CNN performs impressively after retraining on suitable face data, considering human-face-specific patterns may further boost performance.
Abstract
from arXiv · showhide
The Faster R-CNN has recently demonstrated impressive results on various object detection benchmarks. By training a Faster R-CNN model on the large scale WIDER face dataset, we report state-of-the-art results on two widely used face detection benchmarks, FDDB and the recently released IJB-A.
1. Introduction
Region-based CNN methods have become the main paradigm for object detection, with successive generations improving performance and processing speed. This report applies Faster R-CNN to face detection and evaluates it on FDDB and IJB-A.
- Region-based CNN detection methods are now the main paradigm in object detection.
- Three generations of region-based CNN detectors have emerged in recent years with increasingly better performance and faster processing speed.
- The report demonstrates state-of-the-art face detection results using Faster R-CNN on the FDDB and IJB-A benchmarks.
- The report compares multiple generations of region-based CNN detectors and other recent high-performing face detectors.
2. Overview of the Faster R-CNN
Faster R-CNN addresses the computational cost and hand-crafted nature of proposal generation by learning proposals with an RPN and sharing convolutional features with Fast R-CNN. Its pipeline uses anchors to represent objects across scales and aspect ratios.
- R-CNN limitations: R-CNN extracts convolutional features separately for each object proposal, creating a heavy computational burden.
- Proposal generation: R-CNN and Fast R-CNN depend on generic proposals from hand-crafted methods, whose generation can dominate total processing time.EdgeBox proposal generation took 2.73 seconds in the reported experiments.
- Faster R-CNN: Faster R-CNN uses a fully convolutional Regional Proposal Network to generate proposals and a Fast R-CNN detector to refine them.
- Faster R-CNN: Sharing convolutional layers lets the image pass through the CNN once for proposal generation and refinement while supporting deep networks such as VGG16.
- Runtime comparison: On FDDB, the Faster R-CNN pipeline runs significantly faster than both R-CNN and Fast R-CNN.
- RPN design: The RPN uses anchors with 3 scales and 3 aspect ratios, producing k = 9 anchors at each convolutional-map location.For a feature map of size W × H, this yields at most WHk possible proposals.
3. Experiments
Experiments train and evaluate Faster R-CNN face detection using WIDER, FDDB, and IJB-A, comparing proposal generation and region-based CNN detectors. Faster R-CNN achieves strong benchmark performance, with advantages from learned proposals and shared convolutional computation.
- 3.1. Setup: The model is trained on WIDER, which contains 12,880 training images and 159,424 faces with substantial variation in scale, pose, and face count.The dataset's variation makes face detection challenging.
- 3.1. Setup: The experiments test the trained face detector on FDDB and IJB-A, with fine-tuning on IJB-A to accommodate its larger face annotations.WIDER annotations tightly cover facial regions, whereas IJB-A can include areas such as hair.
- 3.2. Comparison of Face Proposals: The RPN and DeepBox achieve higher proposal detection rates than EdgeBox and Faceness, while the RPN performs slightly better than DeepBox.The comparison varies the IoU threshold and measures detection rates among the top N proposals.
- 3.3. Comparison of Region-based CNN Methods: Faster R-CNN significantly outperforms R-CNN and Fast R-CNN on FDDB and also runs faster than both alternatives.The comparison uses Faceness-generated proposals for R-CNN and Fast R-CNN, while Faster R-CNN generates proposals with its RPN.
- 3.4. Comparison with State-of-the-art Methods: For 500 false positives on FDDB, true positive rates are 0.952 with discrete scores and 0.718 with continuous scores.Faster R-CNN performs better than all compared methods for discrete scores beyond around 200 false positives, but is poorer than MultiresHPM under continuous scoring.
- 3.4. Comparison with State-of-the-art Methods: On IJB-A, Faster R-CNN performs better than all compared methods by a large margin, and qualitative results cover overlapping faces and extreme poses and scales.The IJB-A comparison uses results borrowed from prior work.
4. Conclusion
The paper reports state-of-the-art face detection on two benchmarks with Faster R-CNN and attributes its effectiveness to the RPN and shared convolutional layers. The generic detector performs well after retraining on a suitable face dataset, while face-specific patterns may further improve it.
- 4. Conclusion: Faster R-CNN achieves state-of-the-art face detection performance on two benchmark datasets.The conclusion summarizes the reported results across the paper's two evaluated benchmarks.
- 4. Conclusion: The RPN is suggested as the source of effectiveness because it shares convolutional layers with the Fast R-CNN detector without extra computational burden.This sharing permits use of a deep CNN in the RPN without additional computational burden.
- 4. Conclusion: Although designed for generic object detection, Faster R-CNN performs impressively for faces after retraining on a suitable face detection training set.The paper notes that considering special human-face patterns may further boost performance.