Shelf Allocation Learning Data Generation With Realistic Product Layouts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating learning data for shelf allocation analysis often produce composite images that significantly differ from actual product displays, making it difficult to construct an efficient learning model.
Innovation Solution
A learning data generation device that acquires images and information about the arrangement state, deforms images to match the desired arrangement, and generates composite images with similar states, creating learning data that includes these images and their arrangement information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If composite images are generated by synthesizing product images with shelf background images using conventional methods, then learning data can be produced for model training, but the display conditions in the generated images significantly differ from actual product shelf displays, reducing learning effectiveness
Solution Approach 1:
The patent applies preliminary action by extracting arrangement state information from reference images before generating composite images. The system pre-processes reference images to identify object positions, orientations, and spacing relationships, then uses this extracted information to guide the composite image generation process, ensuring that generated images reflect realistic arrangement patterns observed in actual shelf displays
Solution Approach 2:
The patent implements feedback by using extracted arrangement state information as a constraint and reference during composite image generation. The system continuously refers back to the arrangement patterns identified in reference images, adjusting the placement and configuration of products in generated images to match observed real-world arrangement states, thereby improving fidelity
2Productivity
If conventional image synthesis techniques are used to generate learning data, then image generation can be performed quickly, but the generated images do not accurately represent actual product display states, making it difficult to construct efficient learning models
Solution Approach 1:
The system performs preliminary extraction of arrangement state information from reference images before the composite generation process. By pre-identifying object positions, orientations, and spatial relationships in real shelf images, the system establishes accurate reference data that guides subsequent automated generation, ensuring both speed and accuracy
Solution Approach 2:
The patent applies copying by replicating realistic arrangement patterns from reference images into generated composite images. Instead of relying on conventional synthesis that creates unrealistic displays, the system copies authentic spatial relationships, orientations, and configurations from actual shelf images, preserving the reliability of shelf allocation representation while enabling rapid generation of diverse training data
3Adaptability or versatility
If attribute information such as contrast, size, rotation angle and noise are adjusted to generate multiple composite images, then variety in learning data can be increased, but the arrangement state becomes unrealistic and significantly different from actual product displays
Solution Approach 1:
The patent applies local quality by applying attribute adjustments selectively rather than uniformly across all elements. The system maintains accurate arrangement states for critical elements like product positions and orientations while allowing controlled variations in less critical attributes such as lighting conditions, background noise levels, and image contrast, thereby preserving arrangement accuracy while introducing beneficial variety
Solution Approach 2:
The system implements parameter changes by modifying image attributes within constrained ranges that preserve arrangement realism. Instead of allowing arbitrary changes to rotation angles and sizes that would distort arrangement states, the system adjusts parameters like brightness, contrast, and noise levels while maintaining the spatial relationships and orientations extracted from reference images, achieving diversity without sacrificing accuracy
Data Source
AI summary
In a learning data generation device, an acquisition means acquires a first image corresponding to an image obtained by photographing one or more objects arranged in a predetermined arrangement state, and first information corresponding to information indicating the predetermined arrangement state in the first image. An image deformation means deforms a second image corresponding to an image of a desired object. An image generation means generates a fourth image having an arrangement state that matches or is similar to the predetermined arrangement state indicated by the first image, by pasting a deformed second image to the first image or a third image corresponding to the image of a position where the desired object is arranged. A learning data generation means generates data including the fourth image and second information corresponding to the information indicating the arrangement state of the object in the fourth image, as learning data.