Annotation-Efficient Crowd Counting: A Survey with a Focus on Count-Level Weak Supervision

The current crowd counting models are trained from point level head annotation but only the total crowd count is assessed. It is expensive to close this gap: It takes about 2, 000 human-hours to annotate UCF-QNRF. This survey covers the topic of annotation-free/annotation-efficient crowd counting methods particularly with count-level weak supervision, which train from only image-level counts. We categorize the literature by the level of annotation efficiency, and argue that there are a 4-tiered taxonomy and two annotated efficiency pathways—one to enhance the use of annotation, and the other to diminish the need for it. The first pathway includes annotation-efficient losses under point supervision and semi-supervised methods which utilize unlabeled images using uncertainty-aware teacher student training, ranking and surrogate tasks. The second focuses on weakly supervised methods at the count level – from early running ranking networks, MAPS, and MATT, to most recent transformer, MLP and hybrid architectures. On four benchmarks, it is compared with a semi-supervised method and the results demonstrate that count-level weak supervision is competitive to semi-supervision without any semantic location annotations, but with a performance gap to full supervision on dense datasets. With that said, the possibilities of merging distinct approaches to annotation, annotation noise, and foundation models are open challenges.