Crowd Counting dataset preparation is a time-consuming and cumbersome task. The annotations provided in the dataset are manually annotated for each person in the image. Another aspect to crowd data is the dearth of large count data. As a result of this the data distribution in all the crowd counting datasets shows a heavy tailed distribution with large discontinuous gaps towards the tail end. This trend is visually shown for JHU, NWPU , UCF-QNRF, Shanghai-Tech Part A and Part-B in the tabs under the Datasets in Crowd Counting tab.
The approaches used in our work are summarized in the Approaches in the Crowd Counting tab. The approaches are divided into groups based on the type of their input and output. The SOTA tab shows scatter plot performance for each test image for the best performing network for a given dataset. The plots indicate the deviation of the customarily used Mean Absolute Error. An Overall picture for all the networks dataset-wise is provided in the Overall tab. THe deviation for each of the network is almost 4 times the mean deeming the use of MAE unreliable. This motivates the need for alternative training and evaluation methods presented in the Proposed Approach tab.