Generating High-Quality Crowd Density Maps using Contextual Pyramid CNNs

摘要

We present a novel method called Contextual Pyramid CNN (CP-CNN) forgenerating high-quality crowd density and count estimation by explicitlyincorporating global and local contextual information of crowd images. Theproposed CP-CNN consists of four modules: Global Context Estimator (GCE), LocalContext Estimator (LCE), Density Map Estimator (DME) and a Fusion-CNN (F-CNN).GCE is a VGG-16 based CNN that encodes global context and it is trained toclassify input images into different density classes, whereas LCE is anotherCNN that encodes local context information and it is trained to performpatch-wise classification of input images into different density classes. DMEis a multi-column architecture-based CNN that aims to generate high-dimensionalfeature maps from the input image which are fused with the contextualinformation estimated by GCE and LCE using F-CNN. To generate high resolutionand high-quality density maps, F-CNN uses a set of convolutional andfractionally-strided convolutional layers and it is trained along with the DMEin an end-to-end fashion using a combination of adversarial loss andpixel-level Euclidean loss. Extensive experiments on highly challengingdatasets show that the proposed method achieves significant improvements overthe state-of-the-art methods.