Towards using count-level weak supervision for crowd counting

Most existing crowd counting methods require object location-level annotation which is labor-intensive and time-consuming to obtain. In contrast, weaker annotations that only label the total count of objects can be easy obtain in many practical scenarios. This paper focuses on problem weakly-supervised learns a model from small amount (fully-supervised) large count-level (weakly-supervised). Our study reveals most straightforward, is, directly regressing integral density map count, fails provide satisfactory performance. As an alternative solution, we propose method by taking advantage fact estimated via different-but-equivalent maps. key idea enforce consistency between those maps weakly labeled images as regularization terms. We realize this using multiple estimation branches carefully devised asymmetry training strategy, called Multiple Auxiliary Tasks Training (MATT). Through extensive experiments datasets newly proposed dataset, validate effectiveness demonstrate its superior performance over solutions.