Crowd Counting Via Perspective-Guided Fractional-Dilation Convolution

Crowd counting is critical for numerous video surveillance scenarios. One of the main issues in this task how to handle dramatic scale variations pedestrians caused by perspective effect. To address issue, paper proposes a novel convolution neural network-based crowd method, termed Perspective-guided Fractional-Dilation Network (PFDNet). By modeling continuous variations, proposed PFDNet able select proper fractional dilation kernels adapting different spatial locations. It significantly improves flexibility state-of-the-arts that only consider discrete representative scales. In addition, avoiding multi-scale or multi-column architecture used other methods, it computationally more efficient. practice, constructed stacking multiple Convolutions (PFC) on VGG16-BN backbone. introducing generalized operation, PFC can ratios domain under guidance annotations, achieving scales pedestrians. deal with problem unavailable information some cases, we further introduce an effective estimation branch PFDNet, which be trained either supervised weakly-supervised setting once has been pre-trained. Extensive experiments show outperforms state-of-the-art methods ShanghaiTech A, B, WorldExpo'10, UCF-QNRF, UCF_CC_50 and TRANCOS dataset, MAE 53.8, 6.5, 6.8, 84.3, 205.8, 3.06 respectively.