Teaching Data Generation with Orientation for Inclined Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection techniques using neural networks struggle with accurately estimating the orientation of objects within images, leading to significant gaps between bounding boxes and objects, especially when the object's longitudinal direction is inclined, and manually adding orientation to teaching data is burdensome.
Innovation Solution
A teaching data conversion device and method that utilizes a self-learning geometric transformer to automatically calculate and add orientation information to teaching data, including a storage unit for a learned neural network, a calculation unit to determine object orientation, and a generation unit to generate teaching data with object category, position, size, and orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a bounding box is used to represent object position and size, then the object detection can be implemented, but a large gap is formed between the object and the bounding box when the object's longitudinal direction is inclined
Solution Approach 1:
The patent extends the traditional 2D bounding box representation by adding orientation information, effectively moving to a higher dimensional representation. The geometric transformation parameters (including rotation angles) add a new dimension to the bounding box description, allowing the model to represent inclined objects more accurately without increasing the fundamental 2D nature of image processing.
Solution Approach 2:
The patent introduces geometric transformation parameters (scaling, rotation, translation) as additional parameters to the traditional bounding box representation. By changing the parameter set from just position and size to include orientation and transformation metrics, the system can accurately represent inclined objects and reduce the gap between bounding boxes and actual objects.
2Measurement precision
If orientation information is manually added to teaching data, then the object detection accuracy can be improved, but the process becomes burdensome and impractical for large amounts of data
Solution Approach 1:
The patent enables the system to automatically generate and extract orientation information from images through neural network processing, eliminating the need for manual annotation. The model self-learns geometric transformation parameters directly from image data, making the system self-sufficient in generating teaching data with orientation information at scale.
Solution Approach 2:
The patent replaces the manual mechanical process of orientation annotation with an automated neural network-based system. Instead of human operators manually measuring and inputting orientation data, the system uses deep learning models to automatically extract geometric transformation parameters, substituting human labor with computational processing.
3Measurement precision
If the bounding box is rotated to match the object's longitudinal direction, then the alignment improves, but the complexity of the detection system increases
Solution Approach 1:
The patent performs geometric transformation parameter extraction as a preliminary step before final object detection and classification. By pre-processing the images to extract scaling, rotation, and translation parameters, the system prepares the data in advance, making the subsequent detection process simpler and more efficient rather than attempting to solve all problems simultaneously.
Solution Approach 2:
The patent divides the object detection process into separate stages: first extracting geometric transformation parameters (scaling, rotation, translation), then performing the actual detection and classification. This segmentation of the detection pipeline allows each component to focus on a specific task, reducing overall system complexity while maintaining high accuracy.
Data Source
AI summary
The first neural network is a learned neural network learned in such a way as to output, when an object image is input, a geometric transformation parameter relevant to the object image. The object image is an image of an object identified based on object information of the first teaching data including an image and the object information including a category, a position, and a size of an object included in the image. The calculation unit calculates an orientation of the object, based on the geometric transformation parameter being output from the first neural network. The generation unit generates, by adding the orientation of the object being calculated by the calculation unit to the first teaching data, second teaching data including an image and object information including a category, a position, a size, and an orientation of an object included in the image.