Enhancement in Realistic Fluid Re-simulation

Abstract

Abstract-Re-simulating a 3D fluid result from video while retaining and rendering details is significant in practice, which still remains a difficult task in spite of rapid advancements in this field during the last two decades. Physically driven models can be easily extended to handle fluid, yet they are unable to preserve surface details like breaking waves without the expense of increasing particle densities. This paper proposes a hybrid particle Lattice Boltzmann Model for Shallow Waters (LBMSW) coupling method to simulate fluid with finer details real-time from video. To preserve re-simulation surface details, we couple the detail particles with the distribution function of the LBMSW model. To improve the time performance, we use only surface to re-simulate. Experiments show that the proposed approach can obtain a realistic re-simulation products from video example in several challenging scenarios and we provide qualitative evaluation to the method.

Vehicle classification from low-frequency GPS data with recurrent neural networks

“…They also found that aggregating the input features with statistical functions over temporal windows is instrumental in obtaining better results. In [46], the authors perform multi-task learning via deep recurrent neural networks to predict the next position and travel mode in a sequence of GPS data points. They show that using deep learning allows for a significant improvement over traditional "shallow" models, such as Gaussian processes and Hidden Markov Models.…”