Weakly-Supervised Crowd Counting Learns from Sorting Rather Than Locations

Abstract

In crowd counting datasets, the location labels are costly, yet, they are not taken into the evaluation metrics. Besides, existing multi-task approaches employ high-level tasks to improve counting accuracy. This research tendency increases the demand for more annotations. In this paper, we propose a weakly-supervised counting network, which directly regresses the crowd numbers without the location supervision. Moreover, we train the network to count by exploiting the relationship among the images. We propose a soft-label sorting network along with the counting network, which sorts the given images by their crowd numbers. The sorting network drives the shared backbone CNN model to obtain density-sensitive ability explicitly. Therefore, the proposed method improves the counting accuracy by utilizing the information hidden in crowd numbers, rather than learning from extra labels, such as locations and perspectives. We evaluate our proposed method on three crowd counting datasets, and the performance of our method plays favorably against the fully supervised state-of-the-art approaches.

A Skeleton-based Approach for Campus Violence Detection

“…Shallow modeling approaches cannot learn features independently and instead rely on manual methods to extract features that must be fed into a shallow network for classification [14]. Shallow models are best suited for supervised learning, which requires marked data.…”

Fine-grained Domain Adaptive Crowd Counting via Point-derived Segmentation

“…Besides the supervised methods, several approaches focus on relieving the labeling burdensome. They can be broadly categorized into semisupervised methods [5], [40], [55], [68], [91], [93], weaklysupervised methods [83], self-supervised methods [38], [39] and unsupervised methods [14], [60].…”

TransCrowd: Weakly-Supervised Crowd Counting with Transformer

“…1) They apply the point-level annotations to generate ground truth density maps, which are usually expensive cost. Actually, some methods [67,15,40,21] discover that we can collect a new crowd dataset by using a more economical strategy, such as mobile crowd-sensing [15] technology or GPS-less [40] energy-efficient sensing scheduling. For a given crowd scene with different viewpoints and the total count keeps the same (such as auditoria, classroom), if we know the total count of one viewpoint, then the total count of other viewpoints is known.…”

“…Based on the above observations, it is desirable to develop the count-level crowd counting method. Following previous works [21,67], we call the methods which rely on the point-level annotations are fully-supervised paradigm, and the methods which only rely on count-level are weaklysupervised paradigm. The fully-supervised methods first utilize the point annotation to generate the ground truth density map and then elaborately design a regressor to generate a prediction density map and finally apply the L 2 loss to measure the difference between the prediction and the ground truth, as shown in Fig.…”

“…Wang et al [57] introduce a synthetic crowd dataset named GCC, and the model is pre-trained on the GCC dataset and then fine-tuned on real data. One of the most relevant works for our method is [67], which proposing a soft-label network to facilitate the counting task, directly regressing the crowd number without the supervision of location labels. However, [67] is a CNN-based method, which has a limited respective field.…”