Motion Capture Model Training via Quantization Error Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Motion capture technologies using inertial measurement units (IMUs) suffer from accumulating positioning errors, leading to decreased accuracy over time.
Innovation Solution
A method and apparatus for model training in motion capture systems, where a quantization node is introduced into the time series prediction unit to simulate positioning errors, allowing the model to converge towards a quantization-friendly direction and eliminate IMU positioning errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If IMU pose-related information is used for motion capture, then real-time capture capability is achieved, but positioning error accumulates over time
Solution Approach 1:
The patent introduces quantization nodes into the time series prediction unit that deliberately generate quantization errors during training. These errors simulate the harmful IMU positioning errors, allowing the model to learn compensation mechanisms. The harmful quantization errors are converted into a beneficial training signal that enables the model to eliminate real IMU positioning errors during inference.
Solution Approach 2:
The patent implements a feedback mechanism where quantization errors are continuously introduced during training, and the weight parameters are adjusted based on the gradient updates. This closed-loop process ensures the model continuously adapts to and compensates for positioning errors, transforming the error accumulation problem into a learnable pattern that the model can correct.
2Ease of manufacture
If quantization operation is performed on model data, then model data is processed, but positioning error is introduced
Solution Approach 1:
The patent performs quantization operations during the training phase before the model is deployed. By pre-introducing quantization errors and adjusting weight parameters accordingly, the model is prepared in advance to handle positioning errors. This preliminary action ensures that when the model is used for actual motion capture, it can compensate for quantization and positioning errors effectively.
3Reliability
If weight parameter is adjusted according to gradient update, then model converges, but positioning error accumulates
Solution Approach 1:
The patent decomposes the weight parameter into a weight scaling parameter and a weight direction, allowing independent adjustment of each component. The weight scaling parameter is updated based on gradient magnitude while the weight direction is adjusted to maintain quantization friendliness. This parameter decomposition enables the model to converge reliably while simultaneously compensating for positioning errors through optimized parameter updates.
Data Source
AI summary
Embodiments of the present application provide a model training method and apparatus, and a device and a storage medium, and relate to the technical field of motion capture. The method includes: training a motion capture model according to a preset training set, wherein the motion capture model includes a time series prediction unit; performing, through a quantization node in the time series prediction unit, a quantization operation and an inverse quantization operation in sequence on model data passing through the quantization node; and adjusting a weight parameter of the time series prediction unit according to an update on a gradient of the time series prediction unit until the motion capture model converges, wherein the weight parameter includes a weight scaling parameter and a weight direction.