Abstract
Understanding crowd behavior in video is challenging for computer vision. There have been increasing attempts on modeling crowded scenes by introducing ever larger property ontologies (attributes) and annotating ever larger training datasets. However, in contrast to still images, manually annotating video attributes needs to consider spatiotemporal evolution which is inherently much harder and more costly. Critically, the most interesting crowd behaviors captured in surveillance videos (e.g., street fighting, flash mobs) are either rare, thus have few examples for model training, or unseen previously. Existing crowd analysis techniques are not readily scalable to recognize novel (unseen) crowd behaviors. To address this problem, we investigate and develop methods for recognizing visual crowd behavioral attributes without any training samples, i.e., zero-shot learning crowd behavior recognition. To that end, we relax the common assumption that each individual crowd video instance is only associated with a single crowd attribute. Instead, our model learns to jointly recognize multiple crowd behavioral attributes in each video instance by exploring multiattribute cooccurrence as contextual knowledge for optimizing individual crowd attribute recognition. Joint multilabel attribute prediction in zero-shot learning is inherently nontrivial because cooccurrence statistics does not exist for unseen attributes. To solve this problem, we learn to predict cross-attribute cooccurrence from both online text corpus and multilabel annotation of videos with known attributes. Our experiments show that this approach to modeling multiattribute context not only improves zero-shot crowd behavior recognition on the WWW crowd video dataset, but also generalizes to novel behavior (violence) detection cross-domain in the Violence Flow video dataset.
Abstract (translated)
URL
https://arxiv.org/abs/1908.05877
