A ROBUST FACE DETECTION, AGE ESTIMATION AND GENDER PREDICTION BENCHMARK

Abstract

In recent years, there is demand for advanced computer vision technologies particularly in areas like: facial detection and recognition, age estimation, and gender classification. However, the performance of machine learning models in these domains relies heavily on the availability of robust, diverse, and high-quality data. While existing datasets like the COCO, PASCAL VOC, Roboflow100, Adience, UTKFace, and IMDB-WIKI datasets etc., have made significant contributions, still some challenges in terms of demographic diversity, image quality, or annotations accuracy. Moreso, it is crucial to understand that due to image retrieval and annotation costs, most of these datasets consist largely of images found on the web and in some cases do not represent real-life ground truth age labels. As a result, it is difficult to assert with certainty the degree of generalization learned by models on these datasets. To address the aforementioned and support future research, we introduce new dataset, with the goal of advancing the state-of-the-art. Version one of this dataset contains rich human annotations, including face bounding boxes, ground truth age and gender for images of subjects in a natural uncontrolled environment. We obtain a total of 1000 participants videos captured, extracted images from the videos and made available annotations. Moreso, we compare performance using YOLO family state-of-the-art object detector trained on our dataset in comparison to other related benchmarks.