Two-Stage Object Detection Inference With Aggregate and Packed Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning inference processing, particularly object detection tasks, require significant computational resources and time, making it challenging to achieve timely results, especially in emergency scenarios where rapid object detection is necessary.
Innovation Solution
An information processing device that aggregates and packs data to reduce the number of inference regions, employing primary and secondary inferences using trained models to expedite the detection process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of images used for inference is increased to improve detection accuracy, then the accuracy of object detection and recognition is improved, but the inference processing time is extended
Solution Approach 1:
The patent divides the inference process into two distinct stages: a first inference stage that processes a reduced number of aggregate images to identify candidate regions, and a second inference stage that processes only the specific regions of interest identified in the first stage. This segmentation allows the system to maintain high detection accuracy by focusing computational resources on promising areas while avoiding the need to process all images at full resolution, thereby reducing overall inference time.
Solution Approach 2:
The patent performs preliminary processing by generating aggregate images that combine multiple original images into fewer composite images. This preliminary action reduces the total number of images requiring full inference processing. The first inference is then performed on these pre-processed aggregate images to identify candidate regions before conducting the second inference, effectively preparing the data in advance to minimize subsequent computational workload.
2Reliability
If machine learning inference processing is performed on multiple images to improve detection accuracy, then the reliability of detection results is improved, but the calculation load increases
Solution Approach 1:
The patent segments the inference workload into two phases: a first inference phase that operates on a smaller set of aggregate images to identify candidate regions, and a second inference phase that operates only on the specific regions identified as candidates. This segmentation strategy maintains detection reliability by ensuring that multiple images are still considered (through aggregation) while significantly reducing the total calculation load by avoiding full inference on every individual image at high resolution.
Solution Approach 2:
The patent creates aggregate images that are composite representations of multiple original images, serving as simplified copies that capture essential information from multiple sources. These aggregate images can be processed more efficiently than the full set of original images, reducing computational load while preserving the reliability benefits of multi-image analysis.
Data Source
AI summary
An information processing device according to the present invention includes: a memory; and at least one processor coupled to the memory. The processor performs operations. The operations include: generating, based on inference target data include a detection target object, aggregate data smaller than the inference target data; and applying the aggregate data to a trained model to infer a first region that is a region of the detection target object in the aggregate data, generating packed data that is data packed with the first region or a region of the inference target data, applying the packed data to a trained model to infer a second region that is a region of the detection target object that is included, and inferring the detection target object in the inference target data based on the second region.