Human Motion Prediction via Continual Prior Compensation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human Motion Prediction (HMP) involves forecasting future human poses from past motion sequences.
• Existing HMP methods often train predictions for all temporal moments simultaneously, hindering short-term prediction accuracy due to long-term prediction interference.

Purpose of the Study:

• To develop a novel temporal continual learning framework to progressively train HMP models.
• To address the limitation of simultaneous training in HMP by dividing the task into subtasks.
• To mitigate prior information forgetting during progressive training.

Main Methods:

• Introduced Continual Prior Compensation (CPC), a framework that divides HMP into subtasks trained sequentially.
• Developed a learnable Prior Compensation Factor (PCF) to quantify and compensate for prior knowledge loss.
• Enhanced CPC to CPC++ with a Fine-Grained Prior Compensation Factor (FGPCF) for more precise prior loss estimation per subtask.

Main Results:

• CPC and CPC++ frameworks demonstrate effectiveness in improving HMP accuracy.
• The proposed methods are flexible and can be integrated with various HMP backbone models (PGBIG, siMLPe, MotionMixer, LTD).
• Experiments on benchmark datasets validate the superior performance and adaptability of CPC and CPC++.

Conclusions:

• CPC and CPC++ offer a flexible and effective approach to progressive training for Human Motion Prediction.
• These frameworks successfully mitigate the negative impact of long-term predictions on short-term predictions.
• The proposed methods represent a significant advancement in HMP, enhancing accuracy and adaptability across diverse applications.