Self-attention-guided scale-refined detector for pedestrian detection

Abstract

AbstractPedestrian detection has been researched for decades. Recently, an anchor-free method CSP is proposed to generate the pedestrian bounding box directly. When the predicted center deviates from the ground truth in the testing phase, the CSP model generates deviated pedestrian bounding box, which leads to false detection in occlusion situations. To handle this problem, we refine the scale regression branch of the CSP model to generate a more accurate prediction. The new scale regression branch outputs the distances between the center and the four edges of the pedestrian bounding box. Even if the predicted center deviates from the ground truth, an accurate bounding box can still be obtained. Moreover, we integrate a self-attention module into our model to take full advantage of the features in different depth layers. Our proposed model achieves better performance than the state-of-the-art detectors in comparison experiments on the two datasets, i.e., Citypersons and Caltech.