Search

Collective behaviors of coherent groups convey semantic relations among individuals in a crowd scene. Detecting coherent groups is primitive for crowd behavior analysis and practically useful in crowd surveillance. However, classically, crowd analysis in still image is focused on crowd counting estimation or crowd segmentation only. In this paper, we present a novel framework that merges these two classical approaches to achieve a higher level of crowd understanding, i.e., detecting coherent groups within a crowd in a still image. Essentially, in addition to crowd counting estimation, our work can infer crowd segments at both image-level and coherent group level. Extensive experiments and analysis of crowd scene images with a variety of crowd densities demonstrate the efficacy of the proposed framework. The proposed framework also shows its potential application, especially in crowd understanding. [ABSTRACT FROM AUTHOR]

At present, text orientation is not diverse enough in the existing scene text datasets. Specifically, curve-orientated text is largely out-numbered by horizontal and multi-oriented text, hence, it has received minimal attention from the community so far. Motivated by this phenomenon, we collected a new scene text dataset, Total-Text, which emphasized on text orientations diversity. It is the first relatively large scale scene text dataset that features three different text orientations: horizontal, multi-oriented, and curve-oriented. In addition, we also study several other important elements such as the practicality and quality of ground truth, evaluation protocol, and the annotation process. We believe that these elements are as important as the images and ground truth to facilitate a new research direction. Secondly, we propose a new scene text detection model as the baseline for Total-Text, namely Polygon-Faster-RCNN, and demonstrated its ability to detect text of all orientations. Images of Total-Text and its annotation are available at https://github.com/cs-chan/Total-Text-Dataset. [ABSTRACT FROM AUTHOR]

Dense crowd density estimation is one of the fundamental tasks in crowd analysis. While tremendous progress has been made to understand crowd scenes along with the rise of Convolutional Neural Networks (CNNs), research work on dense crowd density estimation is still an ongoing process. In this paper, we propose a novel approach to learn discriminative crowd features from granules, that conforms to the outline between crowd and background (i.e. non-crowd) regions, for density estimation. It shows that by studying the inner statistics of granules for density estimation, this approach is adaptive to arbitrary distribution of crowd (i.e. scene independent). Multiple features fusion is proposed to learn discriminative crowd features from granules. This is to be used as description of the crowd where a direct mapping between the features and crowd density is learned. Extensive experiments on public benchmark datasets demonstrate the effectiveness of our novel approach for scene independent dense crowd density estimation. [ABSTRACT FROM AUTHOR]