Professional: XQLFW
Conference Paper
XQLFW: Cross-Quality Labeled Faces in the Wild. This conference paper was accepted at the 2021 Face and Gesture Recognition Conference (FG 2021) in Jodpur, India.
Face recognition technology frequently encounters issues when dealing with images that aren't pristine. Factors like the distance of the subject, camera settings, or simple blurriness can hinder its effectiveness. Many existing studies have attempted to gauge the impact of these imperfections by artificially adjusting image quality. However, their methods often fall short in replicating the challenges of real-world scenarios. Recognizing this gap, we delved deeper and created a new set of images, the Cross-Quality Labeled Faces in the Wild (XQLFW). Inspired by the well-known Labeled Faces in the Wild (LFW) dataset, our version emphasizes differences in image quality, and crucially, it only incorporates images that have been modified in ways mirroring realistic circumstances.
Face recognition technology frequently encounters issues when dealing with images that aren't pristine. Factors like the distance of the subject, camera settings, or simple blurriness can hinder its effectiveness. Many existing studies have attempted to gauge the impact of these imperfections by artificially adjusting image quality. However, their methods often fall short in replicating the challenges of real-world scenarios. Recognizing this gap, we delved deeper and created a new set of images, the Cross-Quality Labeled Faces in the Wild (XQLFW). Inspired by the well-known Labeled Faces in the Wild (LFW) dataset, our version emphasizes differences in image quality, and crucially, it only incorporates images that have been modified in ways mirroring realistic circumstances.
We then employed the XQLFW dataset to assess the capabilities of several leading face recognition models. Our results painted an intriguing picture: there's a considerable variation in how these models perform across diverse image qualities. This underscores a critical insight - excelling on the standard LFW dataset doesn't guarantee consistent success in real-life situations with fluctuating image standards.
Our new set of pictures, XQLFW, focuses more on real-life image challenges. To boost research and drive improvements in this field, we've made our findings, image set, and tools publicly available. Our goal? To set a new standard benchmark setting for measuring face recognition accuracy in cross-quality scenarios.
Our new set of pictures, XQLFW, focuses more on real-life image challenges. To boost research and drive improvements in this field, we've made our findings, image set, and tools publicly available. Our goal? To set a new standard benchmark setting for measuring face recognition accuracy in cross-quality scenarios.
Technologies