Application of Improved Transformer Based onWeakly Supervised in Crowd Localization.

Aiming to address the issue of complex preprocessing and post-processing required by existing crowd localization methods that employ pseudo bounding boxes or pre-designed localization maps, an end-to-end crowd localization network based on weakly supervised, LocalFormer, is proposed. In the feature extraction stage, a pure Transformer is used as the backbone network, and a global max-pooling operation is performed on the features of each stage to extract more comprehensive details of human heads. In the encoder-decoder stage, the positional information is embedded into the aggregated features as input to the encoder. Each decoder layer uses a set of trainable embeddings as queries, and takes visual features of the last layer of the encoder as keys and values. The decoded features are then used to predict confidence scores. Finally, a binary module is introduced with an adaptive optimized threshold learner to precisely binarize the confidence maps. Extensive experiments on three datasets in different environments show that the proposed method achieves the best positioning performance. [ABSTRACT FROM AUTHOR]