Relative Object Scale Estimation Using Self-Supervised Image Patches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learnable models require extensive training data and manual effort to provide explicit scale indications of objects in images, making it challenging to estimate relative scales in a computationally efficient and unsupervised manner.
Innovation Solution
A computer-implemented method and system that trains a machine learnable model to estimate relative scales of objects in images by using a scale estimator as an addon to a feature extractor, which aggregates feature maps and learns relative scales through spatial scaling of image data with known scale factors, minimizing the mismatch between actual and estimated relative scales without requiring manual ground truth.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with manually annotated scale data is used to train machine learnable models, then the model can provide explicit scale indications, but it requires extensive training data and manual effort
Solution Approach 1:
The system uses self-supervised learning where the model generates its own training signals by predicting scale relationships between image patches. Instead of requiring manual annotations, the model learns from the inherent structure of the data by comparing scaled versions of the same image content, making the system self-sufficient for training purposes
Solution Approach 2:
The approach performs preliminary scaling operations on image patches to create synthetic training examples. By pre-computing scaled versions of images and using them to train the model, the system prepares training data in advance without requiring manual annotation of scale information
2Measurement precision
If supervised learning with annotated scale data is used, then explicit scale indications can be obtained, but the computational complexity increases
Solution Approach 1:
The system uses a lightweight scale estimator that processes only the necessary features from image patches rather than performing exhaustive analysis. By focusing on relative scale estimation between patches rather than absolute scale measurement, the computational burden is reduced while maintaining sufficient accuracy for scene geometry reconstruction
3Measurement precision
If extensive annotated training data is used, then scale estimation accuracy improves, but the ease of training decreases
Solution Approach 1:
The model generates its own training signals by comparing scale relationships between image patches and their scaled versions. This self-supervised approach eliminates the need for manual annotation effort while still providing sufficient training signals for accurate scale estimation
Solution Approach 2:
The system creates synthetic training examples by copying and scaling existing image patches. Instead of requiring new annotated data, the model generates training signals by creating scaled copies of available images, significantly reducing the effort needed to prepare training data
Data Source
AI summary
A system and computer-implemented method for training a machine learnable model to estimate a relative scale of objects in an image. A feature extractor and a scale estimator comprising a machine learnable model part are provided. The feature extractor may be pretrained, while the scale estimator may be trained by the system and method to transform feature maps generated by the feature extractor into relative scale estimates of objects. For that purpose, the scale estimator may be trained on training data in a specific yet non-supervised manner which may not require scale labels. During inference, the scale estimator may be applied to several image patches of an image. The resulting patch-level scale estimates may be combined into a scene geometry map which may be indicative of a geometry of a scene depicted in the image.