Abstract
Currently, face detection approaches focus on facial information by varying specific parameters including pose, occlusion, lighting, background, race, and gender. These studies only utilized the information obtained from low dynamic range images, however, face detection in wide dynamic range (WDR) scenes has received little attention. To our knowledge, there is no publicly available WDR database for face detection research. To facilitate and support future face detection research in the WDR field, we propose the first WDR database for face detection, called WDR FACE, which contains a total of 398 16-bit megapixel grayscale wide dynamic range images collected from 29 subjects. These WDR images (WDRIs) were taken in eight specific WDR scenes. The dynamic range of 90% images surpasses 60,000:1, and that of 70% images exceeds 65,000:1. Furthermore, we show the effect of different face detection procedures on the WDRIs in our database. This is done with 25 different tone mapping operators and five different face detectors. We provide preliminary experimental results of face detection on this unique WDR database.
Introduction
Most modern face detection algorithms are trained and tested on Low Dynamic Range (LDR) images. These standard 8-bit images have a limited ability to capture detail in scenes with high contrast—such as a person standing in front of a bright window or under a harsh spotlight. In these "Wide Dynamic Range" (WDR) scenarios, LDR images often suffer from overexposure (where bright areas become pure white) or underexposure (where dark areas are lost to blackness and noise).
While WDR imaging technology exists to capture these details, the field lacked a standardized, public database to benchmark how well face detectors actually perform when lighting conditions are extreme. WDR images can avoid the pixel saturation that plagues LDR imaging, providing a much richer source of information for computer vision tasks in the wild.
The WDR FACE Database
To bridge this gap, the researchers developed the WDR FACE database, the first of its kind. Using a scientific CMOS (sCMOS) camera, they captured 16-bit grayscale images. Unlike standard 8-bit images which have only 256 levels of brightness, these 16-bit images provide up to 65,536 levels, capturing a massive amount of detail in both highlights and shadows.
Dataset Characteristics:
- Content: 398 16-bit megapixel WDR images.
- Subjects: 29 individuals.
- Dynamic Range: 70% of the images exceed a 65,000:1 contrast ratio, significantly higher than the ~255:1 ratio of standard LDR datasets.
- Resolution: 2048 x 2048 pixels.
- Scene Diversity: Images were captured across eight different scenes, including one artificially lit scene and seven natural scenes with challenging backlighting.
Evaluation Framework
Since most existing face detectors expect standard LDR input, the authors established a rigorous pipeline for evaluation:
- Tone Mapping Benchmarking: The researchers utilized 25 different Tone Mapping Operators (TMOs). TMOs are algorithms that compress WDR data into an LDR format while attempting to preserve structural details.
- Detector Evaluation: Five popular face detection frameworks—Viola-Jones, MMOD, Faster R-CNN, MTCNN, and YOLOv3—were tested across all tone-mapped versions of the dataset.
Key Results and Insights
The study revealed that the choice of TMO is critical. Some TMOs are designed for human aesthetics, while others are better at preserving the structural features required for machine learning models. The operators developed by Gu and WardGlobal were identified as the most "detector-friendly" across the board.
- Top Performer: YOLOv3 achieved the highest overall detection accuracy, peaking at 92.49% when paired with the "WardGlobal" tone mapping operator.
- Deep Learning Resilience: Modern deep learning-based detectors (like YOLOv3 and MTCNN) proved significantly more robust to the artifacts introduced by tone mapping than traditional "hand-crafted" feature methods like Viola-Jones.
- Scene Difficulty: Accuracy varied significantly by environment. In controlled artificial lighting, detectors reached up to 99.1% accuracy. However, in natural scenes with complex backlighting, performance dropped, highlighting the ongoing challenge of real-world WDR detection.
Conclusion and Future Work
The WDR FACE database provides a vital benchmark for the computer vision community. However, the current approach still relies on tone mapping as a "middleman," which inevitably results in some information loss.
Future research directions include developing "WDR-native" face detectors that operate directly on 16-bit data, creating TMOs specifically optimized for machine vision tasks, and expanding the database to include facial identification and verification tasks in WDR environments.
Summary updated