Center and Scale Prediction: Anchor-free Approach for Pedestrian and Face Detection

Object detection traditionally requires sliding-window classifier in modern deep learning based approaches. However, both of these approaches tedious configurations bounding boxes. Generally speaking, single-class object is to tell where the is, and how big it is. Traditional methods combine ”where” ”how” subproblems into a single one through overall judgement various scales In view this, we are interesting whether can be separated two independent subtasks ease problem definition difficulty training. Accordingly, provide new perspective detecting objects approached as high-level semantic feature task. Like edges, corners, blobs other detectors, proposed detector scans for points all over image, which convolution naturally suited. unlike traditional low-level features, goes higher-level abstraction, that looking central there objects, models already capable such abstraction. blob detection, also predict points, straightforward convolution. Therefore, this paper, pedestrian face simplified center scale prediction task convolutions. This way, method enjoys an anchor-free setting, considerably reducing training configuration hyper-parameter optimization. Though structurally simple, presents competitive accuracy on several challenging benchmarks, including detection. Furthermore, cross-dataset evaluation performed, demonstrating superior generalization ability method.