Crowd Counting based on Multi-level Multi-scale Feature

Crowd counting has drawn more and more attention for its significance in reality application. However, it's still a challenging task because of scale variation in images. In this paper, we propose a model to extract and refine features with abundant scale-relevant information, which consists of Multi-layer Multi-scale Feature Extraction Network (MLMS) and Dependency-based Feature Fusion Network (DFF). MLMS plays a role as feature extractor. Three multi-scale feature extraction modules (MSFE) are designed with dilated convolution layers and inserted in different levels of MLMS, which improve the ability for multi-scale feature extraction. DFF plays a role as feature refiner. DFF explores the dependency between hierarchical features. It's the first time in crowd counting to use Long-short term memory (LSTM) to filter information and fuse the features with the assistance of the dependency. Our model provides new ideas for solving scale-relevant problems from two angels: scale feature extraction and fusion. In this way, our model extracts scale-relevant features and refines the features further. Experiments on four challenging datasets ShanghaiTech Part A/B, UCF_QNRF and UCF_CC_50, getting Mean Absolute Error (MAE) 65.3/8.3/113.2/216.3, demonstrate the effectiveness of the proposed model.