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 to obtain in many practical scenarios. This paper focuses on the problem of weakly-supervised crowd counting which learns a model from a small amount of location-level annotations (fully-supervised) and a large amount of count-level annotations (weakly-supervised). Our study reveals that the most straightforward, that is, directly regressing the integral of density map to the object count, fails to provide satisfactory performance. As an alternative solution, we propose a method by taking advantage of the fact that the total count can be estimated via different-but-equivalent density maps. Our key idea is to enforce the consistency between those density maps and total object count on weakly labeled images as regularization terms. We realize this idea by using multiple density map estimation branches and a carefully devised asymmetry training strategy, called Multiple Auxiliary Tasks Training (MATT). Through extensive experiments on existing datasets and a newly proposed dataset, we validate the effectiveness of the proposed weakly-supervised method and demonstrate its superior performance over existing solutions. (C) 2020 Elsevier Ltd. All rights reserved.

Automated summary

This paper focuses on weakly-supervised crowd counting, where a model is learned from a small amount of location-level annotations and a large amount of count-level annotations. The study reveals that directly regressing the integral of density maps to object count is not satisfactory, and proposes a method to enforce consistency between density maps and object count for better performance. Through experiments, the effectiveness of the proposed weakly-supervised method is validated and shown to outperform existing solutions.