A biologically-inspired model for dynamic saliency detection

This paper proposes a biologically inspired model for dynamic saliency detection. Our work simulates the visual processing procedure in primary visual cortex, which differs from previous work on spatio-temporal information extraction and feature integration. We compute the spatio-temporal features by a 3D Gabor filters to simulate the response of neurons to different stimulus of their relative receptive fields. To integrate meaningful features, perceptual grouping is introduced to eliminate distracting features. The facilitative and suppressive interactions among neurons are simulated by convolution and half-wave rectification. Effective spatial and motion features are outputs of the stable responses of neurons after the interaction. Dynamic saliency maps are computed from these features as previous work did. We compare our model with four state-of-the-art dynamic saliency detection models on the public available ASCMN database. Our model achieves higher score for AUC, CC and NSS metric.