Loss Filtering Factor for Crowd Counting

In crowd counting datasets, each person is annotated by a point, typically representing the center of the head. However, due to the dense crowd, variety of scenarios, significant obscuration and low resolution, label noise in the dataset is inevitable and such label noise has a negative impact on the performance of the model. To alleviate the negative effects of label noise, in this paper we propose the Loss Filtering Factor, which can filter out the losses assumed to be caused by label noise during the training process. By doing so, the model can prioritize non-noise data and focus on it during training and predicting. Extensive experimental evaluations have demonstrated that the proposed Loss Filtering Factor consistently improves the performance of all models across all datasets used in the experiments. On average, it leads to a 5.48% improvement in MAE and 6.43% in MSE. Moreover, the proposed approach is universal and it can be easily implemented into any neural network model architecture to improve performance.