2015-cvpr-deeply learned Attributes for crowed Scene understanding

Link to the paper: http://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Shao_Deeply_Learned_Attributes_2015_CVPR_paper.pdf

This article is based on attribute's understanding of the video of crowd scene and uses CNN to learn the characteristics of the properties described.

The main contribution of the article:

1. Build a new large-scale WWW crowd dataset (8,257 scenes, 10,000 videos), and set it to 94 properties

2. Build a CNN model to learn deep Features

Definition of a property

Properties are mainly based on three aspects: Where, who, why

CNN Model

Motion is calculated based on the 2014-cvpr-scene-independent group profiling in crowd paper.

Experiment

The accuracy of appearance is higher than that of motion from the results, and the combination of the two actually has little effect.

2015-cvpr-deeply learned Attributes for crowed Scene understanding