Object Re-identification network based on Dynamic Snake attention module

In the process of object re-identification, there are phenomena such as small target image pixels and complex and variable feature shapes, which often lead to ineffective information being the focus of existing re-identification network features, resulting in low recognition accuracy. This article proposes an attention network based on adaptive snake shaped diffusion convolution to address this issue. Introducing an attention mechanism based on adaptive snake shaped diffusion convolution before and after feature extraction in the backbone network, adjusting the size and shape of the convolution kernel makes it easier for the network to adaptively collect small and irregular features. Experiments have shown that compared with current advanced re-identification methods, this method has achieved superior performance on both Market1501 and JC-1 datasets.