PROJECT TITLE :

Group Sampling for Scale Invariant Face Detection

ABSTRACT:

For the sake of efficiency, detectors that are based on deep learning have a tendency to detect objects of multiple scales within a single input image. Recent works, such as FPN and SSD, generally use feature maps from multiple layers with different spatial resolutions to detect objects at different scales. For instance, high-resolution feature maps are used for detecting small objects. However, our research shows that features extracted from a single layer of the network are sufficient enough to reliably detect objects of any scale. In this paper, we carefully examine the factors affecting detection performance across a wide range of scales. We come to the conclusion that the key is having a balanced set of training samples at various scales, including both positive and negative examples. We propose a method of group sampling in which the anchors are divided into several groups according to the scale, and we make sure that the number of samples that are used for training is the same for all of the groups. The state of the art can be advanced with our method despite the fact that we only use a single layer of FPN for our features. In order to demonstrate that the suggested approach is effective, exhaustive research and testing have been carried out. In addition, we demonstrate that our method is advantageously applicable to other tasks, such as the detection of objects on the COCO dataset, as well as to other detection pipelines, such as YOLOv3, SSD, and R-FCN. When tested on face detection benchmarks such as FDDB and WIDER FACE datasets, our method achieves state-of-the-art results without the use of bells and whistles.