Fast Face Detector Training Using Tailored Views

Kristina Scherbaum, James Petterson, Rogério Schmidt Feris, Volker Blanz, Hans-Peter SeidelICCV 2013 · OpenReview

Abstract

Face detection is an important task in computer vision and often serves as the first step for a variety of applications. State-of-the-art approaches use efficient learning algorithms and train on large amounts of manually labeled imagery. Acquiring appropriate training images, however, is very time-consuming and does not guarantee that the collected training data is representative in terms of data variability. Moreover, available data sets are often acquired under controlled settings, restricting, for example, scene illumination or 3D head pose to a narrow range. This paper takes a look into the automated generation of adaptive training samples from a 3D morphable face model. Using statistical insights, the tailored training data guarantees full data variability and is enriched by arbitrary facial attributes such as age or body weight. Moreover, it can automatically adapt to environmental constraints, such as illumination or viewing angle of recorded video footage from surveillance cameras. We use the tailored imagery to train a new many-core implementation of Viola Jones' AdaBoost object detection framework. The new implementation is not only faster but also enables the use of multiple feature channels such as color features at training time. In our experiments we trained seven view-dependent face detectors and evaluate these on the Face Detection Data Set and Benchmark (FDDB). Our experiments show that the use of tailored training imagery outperforms state-of-the-art approaches on this challenging dataset.

Introduction

Traditional face detection algorithms, like the classic Viola-Jones framework, rely on massive datasets of manually labeled images. This approach suffers from several critical flaws: the labeling bottleneck where manual annotation is time-consuming and error-prone; a lack of variability in real-world datasets regarding lighting, age, and pose; and the "one-size-fits-all" trap where general detectors struggle in specific scenarios, such as high-mounted surveillance angles. This research addresses these issues by shifting from data collection to automated data generation, creating perfectly labeled, highly variable, and environment-specific training sets.

Methods and Data Generation

The core of this approach is a pipeline for generating "tailored" synthetic training samples using a statistical 3D face model. By modulating Principal Component Analysis (PCA) components, the system generates an infinite variety of unique facial shapes and textures. Beyond basic geometry, "attribute vectors" control semantic features like age, body weight, and skin tone.

View-dependent face detectors ranging from -30° to +30°

Faces are rendered into 2D images with specific parameters for 3D pose, camera focal length, and lighting. Specifically, seven distinct training sets were generated corresponding to viewing angles from -30° to +30°. These rendered faces are then blended onto real-world background scenes from the PASCAL VOC dataset to create positive training samples. The training utilized a many-core implementation of the Viola-Jones AdaBoost framework, extended to support multiple feature channels, including color (RGB).

Key Results and Evaluation

The primary benchmark used was the Face Detection Data Set and Benchmark (FDDB), containing 5,171 annotated faces. A key insight is that training on just 75 synthetic subjects with high statistical variability is enough to outperform detectors trained on much larger sets of real-world images.

ROC curve comparison on the FDDB benchmark

As shown in the ROC curve, the "Tailored View Detector" outperformed several well-known methods, including the original Viola-Jones and various published commercial and academic detectors. Furthermore, the many-core implementation achieved a 5.3x training speedup using eight CPU cores. The inclusion of color channels proved beneficial, with the red channel being the most descriptive for identifying skin tones.

Conclusion

This work demonstrates that synthetic data can effectively replace manual labeling for training robust face detectors. By "tailoring" training imagery to specific environmental constraints, performance is significantly improved. Future work will explore more sophisticated ensemble methods for merging view-specific detectors and investigate alternative color spaces like YCbCr or Lab for enhanced skin-color modeling.

Summary updated

Abstract

Face detection is an important task in computer vision and often serves as the first step for a variety of applications. State-of-the-art approaches use efficient learning algorithms and train on large amounts of manually labeled imagery. Acquiring appropriate training images, however, is very time-consuming and does not guarantee that the collected training data is representative in terms of data variability. Moreover, available data sets are often acquired under controlled settings, restricting, for example, scene illumination or 3D head pose to a narrow range. This paper takes a look into the automated generation of adaptive training samples from a 3D morphable face model. Using statistical insights, the tailored training data guarantees full data variability and is enriched by arbitrary facial attributes such as age or body weight. Moreover, it can automatically adapt to environmental constraints, such as illumination or viewing angle of recorded video footage from surveillance cameras. We use the tailored imagery to train a new many-core implementation of Viola Jones' AdaBoost object detection framework. The new implementation is not only faster but also enables the use of multiple feature channels such as color features at training time. In our experiments we trained seven view-dependent face detectors and evaluate these on the Face Detection Data Set and Benchmark (FDDB). Our experiments show that the use of tailored training imagery outperforms state-of-the-art approaches on this challenging dataset.
BibTeX
@inproceedings{scherbaum2013fast,
  author = {Kristina Scherbaum and James Petterson and Rogério Schmidt Feris and Volker Blanz and Hans-Peter Seidel},
  title = {{Fast Face Detector Training Using Tailored Views}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2013},
  url = {https://openreview.net/forum?id=Jyx1B5x7eNo},
}
Figure