Unsupervised Angle Estimation via Feature Rigid Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training angle estimators require supervision in the form of angle labels or structural data, which is undesirable for unsupervised training.
Innovation Solution
A training apparatus and method that utilize feature extraction, angle estimation, rigid transformation, matching loss computation, and model updating to train angle estimators without any annotation, allowing the features to appear as if they were extracted from images at the same angle.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ground truth angle labels or structural data are used for training, then the accuracy of the angle estimator is improved, but the complexity of data preparation and annotation increases
Solution Approach 1:
The training method enables the system to train angle estimators using only input images without requiring external angle labels or structural data annotations. The system serves itself by computing matching losses between transformed features from different images to automatically learn angle representations, eliminating the need for manual annotation processes while maintaining training effectiveness
2Reliability
If supervision from angle labels or structural data is required, then the training process becomes more guided and convergent, but the applicability to unannotated data decreases
Solution Approach 1:
The method introduces transformed features as an intermediary between input images and angle estimation. By transforming features from one image to another's coordinate system and computing matching losses, the system creates a self-supervised learning signal that guides convergence without requiring external annotations, thereby enabling both reliable training and broad applicability to unannotated data
Data Source
AI summary
The training apparatus includes an angle difference computation section which calculates a difference between angles estimated by one or more angle estimators, a rigid transformation section which transforms the feature of the input image according to the difference, a matching loss computation section which calculates a matching loss between a non-transformed feature extracted by the one or more feature extractors and the feature transformed by the rigid transformation section, and an updating section which updates at least the one or more angle estimators with reference to the matching loss, wherein the rigid transformation section transforms the feature in a way that the feature appears as if it has been extracted from an image at the same angle of the image from which the non-transformed feature value has been extracted.