The paper introduces a novel crowd counting model called Switching Convolutional Neural Network (Switch-CNN) that leverages intra-image crowd density variation to improve the accuracy and localization of predicted crowd counts. The model consists of multiple independent CNN regressors with different receptive fields, each trained to capture specific crowd characteristics. A switch classifier is used to relay patches from a crowd scene to the most suitable regressor based on the patch's attributes. The training process involves differential training, where each regressor is trained on patches that minimize count error, followed by coupled training to co-adapt the switch classifier and regressors. Extensive experiments on major crowd counting datasets (ShanghaiTech, UCF_CC_50, UCSD, and WorldExpo'10) demonstrate that Switch-CNN outperforms state-of-the-art methods in terms of Mean Absolute Error (MAE) and Mean Squared Error (MSE). The model's ability to learn a multichotomy of crowd scene patches based on density is also analyzed, showing that it effectively groups patches based on latent factors correlated with crowd density.The paper introduces a novel crowd counting model called Switching Convolutional Neural Network (Switch-CNN) that leverages intra-image crowd density variation to improve the accuracy and localization of predicted crowd counts. The model consists of multiple independent CNN regressors with different receptive fields, each trained to capture specific crowd characteristics. A switch classifier is used to relay patches from a crowd scene to the most suitable regressor based on the patch's attributes. The training process involves differential training, where each regressor is trained on patches that minimize count error, followed by coupled training to co-adapt the switch classifier and regressors. Extensive experiments on major crowd counting datasets (ShanghaiTech, UCF_CC_50, UCSD, and WorldExpo'10) demonstrate that Switch-CNN outperforms state-of-the-art methods in terms of Mean Absolute Error (MAE) and Mean Squared Error (MSE). The model's ability to learn a multichotomy of crowd scene patches based on density is also analyzed, showing that it effectively groups patches based on latent factors correlated with crowd density.