Multitask Non-Autoregressive Model for Human Motion Prediction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human motion prediction is crucial for understanding human actions.
• Existing RNN-based encoder-decoder models face error accumulation due to autoregressive generation.
• Autoregressive methods generate future poses conditioned on previously predicted ones, leading to cumulative errors.

Purpose of the Study:

• To propose a novel Non-Autoregressive model (NAT) for human motion prediction.
• To address and mitigate the error accumulation problem in human motion prediction.
• To improve the accuracy and efficiency of predicting future human skeletons.

Main Methods:

• Developed a Non-Autoregressive model (NAT) with a non-autoregressive decoding scheme.
• Incorporated a context encoder for temporal and spatial embedding of input poses.
• Utilized a positional encoding module to inject temporal order information.
• Implemented a multitask training paradigm combining skeleton prediction and action recognition.

Main Results:

• The proposed NAT model effectively predicts future human poses without autoregression.
• Achieved state-of-the-art performance on Human3.6M and CMU-Mocap benchmarks.
• Demonstrated significant improvements over existing autoregressive methods in human motion prediction.

Conclusions:

• Non-autoregressive decoding is a viable and effective approach for human motion prediction.
• The NAT model offers a robust solution to error accumulation in sequence generation tasks.
• Multitask learning enhances the performance of human skeleton prediction models.