Monocular Camera Distance Prediction Using Patch-Image Model Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for obtaining distance to a subject using a monocular camera require a large dataset for training, which is difficult to prepare, and have high processing loads due to the need for extensive computation in patch-based methods, while image-based methods struggle with domain adaptation and data set preparation.
Innovation Solution
A learning method that uses a patch method model for preliminary learning to generate a statistical model, which is then used to learn an image-based method model without the need for a separate data set, allowing for efficient prediction of distances using an image-based approach with reduced computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a massive amount of data set is used to generate a high-accuracy statistical model, then the measurement precision of distance is improved, but the ease of manufacture (data set preparation) deteriorates
Solution Approach 1:
The patent uses a synthesized data set generated by a rendering apparatus to copy and replicate training data, eliminating the need to collect real-world images with known distances. The rendering apparatus generates synthetic images with ground truth distance information by computationally rendering scenes with known geometric parameters, thus creating training data without manual measurement or annotation of physical objects.
Solution Approach 2:
The system performs self-service by automatically generating its own training data through the rendering apparatus. Instead of requiring external data collection and manual annotation processes, the statistical model training is supported by self-generated synthetic data sets that contain both image data and corresponding ground truth distance values, making the system self-sufficient in data preparation.
2Measurement precision
If patch-based methods are used to obtain distance, then the measurement precision is improved, but the processing load increases
Solution Approach 1:
The patent replaces the mechanical/computational process of dividing images into patches with a direct image-based approach. Instead of manually or algorithmically segmenting images into smaller patches for processing, the system uses the complete image directly as input to the statistical model, substituting the complex patch-processing mechanism with a simpler whole-image processing approach that maintains precision while reducing computational burden.
3Productivity
If image-based methods are used to obtain distance, then the processing load is reduced, but the adaptability to different domains deteriorates
Solution Approach 1:
The patent addresses domain adaptation by changing the parameters and characteristics of the synthetic data generation process. The rendering apparatus can generate synthetic images with varying camera parameters, scene geometries, object types, and environmental conditions, allowing the statistical model to be trained on diverse data distributions. This enables the model to adapt to different domains and applications without requiring retraining on domain-specific real-world data.
Data Source
AI summary
According to one embodiment, a learning method for causing a second statistical model to learn using a first statistical model is provided. The method includes obtaining a first learning image, cutting out each local area of the obtained first learning image, and obtaining a first prediction value output from the first statistical model by inputting each local area to the first statistical model and obtaining a second prediction value output from the second statistical model by inputting the entire area of the first learning image to the second statistical model, and causing the second statistical model to learn based on a difference between the first prediction value and the second prediction value.