Advancements in Video Prediction with Artificial Intelligence

2019-11-09

Video prediction is a form of unsupervised learning, as it requires predicting every pixel from a sequence of images to learn the unannotated structure of video data. To achieve good results, complex loss functions are used. Additionally, video prediction is a multimodal problem, as accurate predictions require latent information that is not provided in the given videos. Traditional methods are generally ineffective for high-resolution videos, as they struggle to handle the complexity of the real world. However, artificial intelligence and machine learning algorithms have become increasingly proficient in video prediction, accurately predicting, for example, where a baseball might fly. To further enhance this predictive capability, researchers from the University of Michigan, Google, and Adobe have proposed a novel approach. They have developed a large-scale model that can generate high-quality videos using only a few frames, thereby advancing the technology. Unlike traditional methods, this new approach does not rely on techniques such as optical flow (apparent motion patterns of objects, surfaces, or edges in a scene) or landmarks. The researchers stated in a preprint paper describing their work, "To the best of our knowledge, maximizing the capacity of standard neural networks has not been thoroughly investigated for improving prediction accuracy. Our study represents the first comprehensive investigation into the impact of video prediction capacity growth." The team's baseline model is built upon existing architectures for random video generation, with components that model the inherent uncertainty of future predictions. The researchers trained and tested several versions of the model using datasets customized for three prediction categories: object interactions, structured motion, and partial observability. First, the researchers selected 256 videos from a video database and extracted video segments. Then, they adjusted the input conditions for each model to include two to five video frames and trained the models to predict the next five to ten frames during the training process. The models were capable of generating up to 25 frames for all tasks at a low resolution of 64×64 pixels. The researchers reported that the model clearly depicted human arms and legs and made accurate predictions. Furthermore, these predictions appeared very realistic. In conclusion, the authors of the paper wrote, "Our research confirms the importance of recurrent connections and random modeling in situations of uncertainty, such as videos with unknown actions or controls. We also found that maximizing the capacity of these models can improve the quality of video prediction. We hope that our work will advance this field and lead to further developments in similar directions, achieving high-quality video prediction through continuously improving techniques."

RECENT AI NEWS

RECENT AI TOOLS