Self-Conditional Crowd Activity Detection Network with Multi-label Classification Head

Abstract

In this paper, we proposed new head network architecture in deep neural networks to classify categories for crowd activity. The proposed network was motivated by multi-label classification and conditional generative adversarial networks. In the head network, latent features were transformed into multi-label embedding vectors using pre-trained deep neural networks. The multi-label embedding vectors were regarded as the probability of relevant objects' existence. Then irrelevant embedding components were eliminated by the threshold layer. The refined multi-label embedding vectors are combined with pure latent feature vectors. Finally, a last linear layer predicted crowd activities. The proposed models configured ResNet back-bones with pre-trained weights. In terms of mean accuracy performances, our proposed models showed 1.55% higher in the best case, whereas 0.38% less in the worst case by comparing with baseline models.