Object Pose Estimation Training Using Polarization and Distance Cues
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for object pose estimation rely heavily on costly ground truth annotations and struggle with photometrically-challenging objects like reflective surfaces, as they fail to accurately estimate object poses due to unreliable distance and polarization information.
Innovation Solution
A self-supervised training method that leverages distance and polarization information as pseudo-labels to correct and update the pose predictions of machine learning models, using a combination of polarization mask, surface normal, and distance losses to enhance the model's robustness and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ground truth poses are used for training machine learning models, then model training accuracy is improved, but annotation cost increases significantly
Solution Approach 1:
The system uses the machine learning model itself to generate pose predictions that serve as training labels, eliminating the need for manual ground truth annotations. The model predicts poses from images, and these predictions are used to update the model through self-supervised learning, creating a self-service training mechanism.
Solution Approach 2:
The system creates synthetic training data by generating pseudo-labels from the model's own predictions and augmenting them with polarization and distance information. This copying approach allows the model to learn from synthesized data that mimics real ground truth without requiring actual manual annotations.
2Productivity
If traditional training methods are used for photometrically-challenging objects, then training speed is maintained, but pose estimation accuracy deteriorates due to unreliable distance and polarization information
Solution Approach 1:
The system changes the training parameters by incorporating polarization information and distance data into the loss function. This modifies how the model learns from photometrically-challenging objects, using additional physical parameters (polarization angles, distance measurements) to compensate for unreliable visual information from reflective surfaces.
Solution Approach 2:
The system introduces polarization information and distance data as intermediary elements that mediate between the image input and pose output. These intermediaries provide additional constraints and information that help resolve ambiguities in photometrically-challenging scenarios, improving accuracy without sacrificing training speed.
3Measurement precision
If stereo images are used to improve pose estimation, then depth accuracy is improved, but device complexity and processing requirements increase
Solution Approach 1:
The system creates a virtual depth map by synthesizing it from polarization information and the model's pose predictions, rather than using actual stereo image processing. This copied depth information is sufficient for training purposes and avoids the complexity of stereo vision systems.
Solution Approach 2:
The system extracts only the essential depth information needed for training by deriving it from polarization data and pose predictions, rather than processing full stereo image pairs. This extraction approach obtains necessary depth accuracy while avoiding the computational burden of complete stereo processing.
Data Source
Figure 1
Figure 2(a)~3(b)
AI summary
A computer-implemented training method for a machine learning model configured to estimate object pose in an image (20), the computer-implemented training method comprising: - obtaining at least one image (20) of an object taken from a viewpoint, distance information (30) between the object and the viewpoint and polarization information (40) of the object, and, for each of the at least one image: - predicting a pose (24) of the object by the machine learning model based on said image (20), and - updating the machine learning model based on the predicted pose (24), the distance information (30) and the polarization information (40).