A data-driven approach to cleaning large face datasets

Abstract

Large face datasets are important for advancing face recognition research, but they are tedious to build, because a lot of work has to go into cleaning the huge amount of raw data. To facilitate this task, we describe an approach to building face datasets that starts with detecting faces in images returned from searches for public figures on the Internet, followed by discarding those not belonging to each queried person.We formulate the problem of identifying the faces to be removed as a quadratic programming problem, which exploits the observations that faces of the same person should look similar, have the same gender, and normally appear at most once per image. Our results show that this method can reliably clean a large dataset, leading to a considerable reduction in the work needed to build it. Finally, we are releasing the FaceScrub dataset that was created using this approach. It consists of 141,130 faces of 695 public figures and can be obtained from http://vintage.winklerbros.net/facescrub.html.

Labeled Faces in the Wild: A Survey

“…The FaceScrub data set [71] contains 107, 818 images of celebrities automatically collected from the web, and verified using a semi-automated process. It contains 530 different individuals, with an average of approximately 200 images per person.…”

A Survey of the Trends in Facial and Expression Recognition Databases and Methods

“…This caused a shift of focus to methods that are independent of image storage and compression limitations. Some of the well-known recent databases collect images from videos in the internet and they represent a wide variation in image storage and quality [15] [25][26][27][28]. This motivates further research into scalable cloud-based methods that can extract features from large databases and correlate them with facial recognition tasks.…”

“…From the year 2000 and onwards, the facial databases were seen to capture the variations in pose, lighting, imaging angles, ethnicity, gender and facial expressions [4]. Some of the most recent databases capture the variations in image sizes, compression, occlusions and are gathered from varied sources such as social media and internet [15].…”

Biologically Inspired Vision for Human-Robot Interaction

“…To detect a face/hand we employ a modified HOG (Histogram of Oriented Gradients) descriptor combined with responses of complex cells and a linear SVM to code the shape. The face and hand detectors were trained and evaluated on the FaceScrub dataset [13] and the Oxford hand dataset [14], respectively. The developed HRI system does not need any prior calibration and has been designed to run in real time.…”

“…Finally, non-maximum suppresion is applied to eliminate multiple detections of the same face/hand (see below). We used the FaceScrub dataset [13] and the Oxford Hand dataset [14] to train and evaluate our face and hand detectors. For each detector we train an initial classifier using the positive and a random set of negative examples, then we use it to scan over images not containing faces or hands and collect false positives, and then we do a second round of training by including these hard false positives into the negative training set.…”