Composition Loss for Counting, Density Map Estimation and Localization in Dense Crowds

Abstract

With multiple crowd gatherings of millions of people every year in events ranging from pilgrimages to protests, concerts to marathons, and festivals to funerals; visual crowd analysis is emerging as a new frontier in computer vision. In particular, counting in highly dense crowds is a challenging problem with far-reaching applicability in crowd safety and management, as well as gauging political significance of protests and demonstrations. In this paper, we propose a novel approach that simultaneously solves the problems of counting, density map estimation and localization of people in a given dense crowd image. Our formulation is based on an important observation that the three problems are inherently related to each other making the loss function for optimizing a deep CNN decomposable. Since localization requires high-quality images and annotations, we introduce UCF-QNRF dataset that overcomes the shortcomings of previous datasets, and contains 1.25 million humans manually marked with dot annotations. Finally, we present evaluation measures and comparison with recent deep CNN networks, including those developed specifically for crowd counting. Our approach significantly outperforms state-of-the-art on the new dataset, which is the most challenging dataset with the largest number of crowd annotations in the most diverse set of scenes.

From Open Set to Closed Set: Counting Objects by Spatial Divide-and-Conquer

“…We follow the same data augmentation used in [21], except for the UCF-QNRF dataset [15] where we adopt two data augmentation strategies. In particular, 9 sub-images of 1 4 resolution are cropped from the original image.…”

ZoomCount: A Zooming Mechanism for Crowd Counting in Static Images

“…The dataset was recently published by Idrees et al [22], which is a challenging and the first dataset of its kind. On one hand, it contains images with resolution as high as (6666 × 9999) and as low as (300 × 377); on the other hand, crowd count per image ranges from a maximum value of 12, 865 to a minimum count of 65.…”

“…8. We also compare the crowd estimate of ten test images each, for both extreme cases with DenseNet [20] direct regression and the state-of-the-art CL [22] density map method. Our method performs much better in both cases, as shown in Fig.…”

“…Detection-based methods perform well in low crowd images but could not be generalized well to high-density crowd images as detection fails miserably in such cases due to very few pixels per head or person. Density map estimation methods [22], [39], [40] Left and right graphs compare ten cases each, belonging to high dense and low dense extreme respectively for Density map [3], DenseNet [20] based direct regression and our method. As shown, other models either highly underestimate or overestimate, whereas the proposed method remains the closest to the ground truth (GT) bar in most cases.…”

Focus on Semantic Consistency for Cross-Domain Crowd Understanding

“…At present, many CNN-based approaches [4][5][6][7][8] have achieved phenomenal performance, and their success is driven by the availability of public crowd datasets. Unfortunately, the existing datasets (such as Shanghai Tech A/B [9], and UCF-QNRF [10], etc.) are so small-scale that makes it difficult for trained models to perform well in other scenarios.…”