PFIDT:Pyramid Focal Inverse Distance Transformer for Crowd Localization

Precise individual localization in densely crowded scenes remains a formidable challenge due to the complex interplay of occlusions and variable density patterns. Traditional approaches to crowd localization often employ CNNs to regress density maps, yet these are prone to inaccuracies due to the severe overlaps characteristic of dense populations. Addressing this conundrum, this paper introduces the Pyramid Focal Inverse Distance Transformer (PFIDT), a framework that leverages the multi-scale global receptive fields of Pyramid Vision Transformers to effectively conform to the multi-scale nature of crowds, thereby enhancing the precision of individual localization. The PFIDT incorporates Focal Inverse Distance techniques to adeptly tackle issues of scale variation and dense overlaps, prioritizing local small-scale features within the contextual understanding of the scene. Furthermore, the model introduces an independent Structural Similarity Index (I-SSIM) loss, encouraging the learning of local structural information and better recognition of local maxima. Rigorous testing on standard benchmarks, including the challenging ShanghaiTech dataset, has confirmed the supremacy of our method. The PFIDT has achieved outstanding results on the ShanghaiTech dataset, with Mean Absolute Errors (MAE) of 53.4 and 6.5 and Mean Squared Errors (MSE) of 93.5 and 10.7 for Parts A and B.