Two-Stage Inference for Object Detection Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection tasks using machine learning, particularly deep-learning, face challenges in improving the throughput of inference, leading to inefficiencies in processing large volumes of data and potentially reducing accuracy due to high computational loads.
Innovation Solution
An information processing device and method that employs a two-stage inference approach, where a primary inference unit processes data using a first learned model, generates aggregated data with reduced quantity, and then applies a correspondence relation to a secondary inference unit using a second learned model to enhance inference efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single learned model is used for object detection, then the inference process is simple, but the throughput is insufficient when processing large volumes of data
Solution Approach 1:
The inference process is divided into two independent stages: primary inference using a first learned model and secondary inference using a second learned model. Each stage processes data independently, allowing parallel execution and improving overall throughput without creating complex interdependencies between processing steps.
Solution Approach 2:
The primary inference is performed as a preliminary action before secondary inference. The first learned model pre-processes the input data and generates preliminary detection results, which then serve as input for the second learned model. This preliminary processing reduces the computational burden on the secondary inference stage.
2Quantity of substance
If a first learned model is used for primary inference, then preliminary detection results are obtained, but the data quantity remains large requiring further processing
Solution Approach 1:
The second learned model extracts only the necessary target objects from the large volume of preliminary detection results produced by the first learned model. This extraction process filters out irrelevant data and retains only the critical information needed for final detection, significantly reducing the data quantity that requires detailed processing.
Solution Approach 2:
The first learned model acts as an intermediary that transforms the original input data into preliminary detection results, which then serve as input for the second learned model. This intermediary processing step simplifies the data structure and reduces complexity for the final inference stage.
3Measurement precision
If a second learned model is used for secondary inference, then inference accuracy is maintained, but the computational load increases
Solution Approach 1:
The second learned model performs partial inference by focusing only on the specific target objects identified by the first learned model, rather than processing all possible objects in the input data. This partial action approach maintains inference accuracy for critical targets while reducing overall computational load.
Solution Approach 2:
The system uses different learned models with optimized parameters for different inference stages. The first learned model is optimized for rapid preliminary detection, while the second learned model is optimized for accurate final detection. This parameter optimization allows each stage to operate efficiently with appropriate computational resources.
Data Source
AI summary
An information processing device includes: a memory and, at least one processor coupled to the memory. The processor performs operations. The operations includes: applying inference target data in which at least part of data includes a first target object to a first learned model to infer the first target object as primary inference; generating aggregated data that is data having a smaller quantity than the inference target data by using the first target object inferred in the primary inference; generating a correspondence relation between a position of the first target object in the inference target data and a position of the first target object in the aggregated data; applying the aggregated data to a second learned model to infer the first target object as secondary inference; and inferring the first target object by using the first target object in a result of the secondary inference and the correspondence relation.