Wide Angle Target Detection via Equirectangular Projection Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques fail to accurately detect targets in wide angle view images, such as spherical content, due to projection transformations like equirectangular projection, which distort features and complicate target detection processes.
Innovation Solution
An image processing apparatus that performs projection transformation on image data to create learning data corresponding to the target's projection scheme, allowing for the generation of a learned model that can accurately detect targets within wide angle view images without the need for additional perspective projection transformations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a learned model corresponding to two-dimensional images obtained by typical perspective projection is used for target detection, then the model can accurately detect targets in standard images, but the target detection accuracy deteriorates in wide angle view images due to projection transformation distortion
Solution Approach 1:
The patent changes the projection parameter from typical perspective projection to equirectangular projection by transforming the learning data. This allows the learned model to adapt to wide angle view images while maintaining detection accuracy, resolving the contradiction between detection precision and adaptability to different image types
Solution Approach 2:
The patent performs projection transformation on the learning data in advance before training the model. By pre-transforming the learning data into equirectangular projection format, the model learns to recognize targets in wide angle view images directly, eliminating the need for runtime perspective projection transformations and improving both accuracy and efficiency
2Ease of operation
If perspective projection transformation is performed on wide angle view images before target detection, then the image format is converted to standard two-dimensional format, but additional processing steps and computational load are required
Solution Approach 1:
Instead of transforming wide angle view images to perspective projection format before detection, the patent inverts the approach by transforming the learning data to equirectangular projection format. This allows direct detection on wide angle view images without requiring perspective projection transformation, simplifying the processing pipeline while maintaining compatibility with learned models
3Productivity
If conventional target detection techniques are applied to wide angle view images, then the detection process can be performed on available images, but target detection accuracy deteriorates due to feature amount differences caused by projection transformation
Solution Approach 1:
The patent changes the projection parameter of the learning data from perspective projection to equirectangular projection, matching the wide angle view image format. This allows the model to learn features specific to equirectangular projection, maintaining both high detection accuracy and processing efficiency when applied to wide angle view images
Data Source
AI summary
An image processing apparatus (100) according to the present disclosure includes: a learning-data creation unit (132) configured to perform projection transformation on image data including a target as a subject, the learning-data creation unit (132) being configured to create learning data including the target as correct data; and a model generation unit (133) configured to generate, based on the learning data created by the learning-data creation unit (132), a learned model for detecting the target included in input data that includes a wide angle view image and is input to the learned model, the wide angle view image being created by projection transformation identical in scheme to the projection transformation by which the learning data is created.