This document proposes a novel method for video saliency detection based on an adaptive nonlinear partial differential equation (PDE) model. The key contributions are:
1. Refining an existing PDE-based static saliency detection model (LESD) to incorporate orientation and motion information important for video saliency detection.
2. Combining static saliency maps generated from the PDE model with motion maps extracted from motion vectors to produce the final saliency map.
3. Extending the model to account for flow-like structures by adding a non-linear matrix tensor to rotate the PDE flow towards orientations of interesting features.