A Densenet Based Robust Face Detection Framework

Abhilash Nandy; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

Face Detection has become important in various real-life applications such as face recognition, kinship verification, video surveillance, sentiment analysis using videos, etc. There has been significant progress in this field in recent years, thanks to the evolution of deep convolutional neural networks (CNNs). Images taken in real-world scenarios vary a lot in various aspects such as lighting, scale, pose, etc. WIDER FACE dataset contains such images, and is hence, quite challenging. In this paper, we propose a solution which takes the DSFD (Dual Shot Face Detector) as a baseline network, and we apply some tweaks to the network to improve performance with lesser memory usage and inference time. Specifically, we use a Densenet backbone, use focal loss function for classification, a function of IoU (Intersection over Union) metric as a regression loss function, and lastly, use the max-out operation before predicting class probabilities. Consequently, the proposed solution achieves state-of-the-art performance on the WIDER FACE Dataset, with added advantages of being more scalable and taking lesser time to infer than its original DSFD baseline. Also, it gives better face detection performance than many other state-of-the-art face detection frameworks.