Crowd Counting with Latent Diffusion Model

Due to their two-stage generation mode, latent diffusion models feature lower computational complexity and are widely applied in image generation. To extend their application scope, we apply LDM(Latent Diffusion Model) to crowd counting domain, generating density maps from crowd images through latent space diffusion denoising sampling. We propose an autoencoder to enable compression and reconstruction of density maps. We propose a guidance mechanism that uses multi-level conditional feature maps for crowd counting task requirements. This mechanism is designed to respectively inject high-level and low-level semantic feature of given crowd image into different stages during the denoising diffusion process. Experiments conducted on public datasets show effectiveness of our proposed method.