UCF-CC-50 is a dataset for counting the number of people in dense crowds in images. The dataset fuses information from three aspects and imposes smoothness constraints on nearby patches to improve the estimation of incorrect patches, thereby producing a better estimate of the number of people in the image.

The dataset includes:

• 50 images
• 63,974 head center annotations (between 94 and 4,543 people per image)
• DPM deformable parts model trained on the INRIA character model

The images are mainly collected from FLICKR (an American image service website).