Objectness Learning Using Overlap Scores for Bounding Box Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using convolutional neural networks (CNNs) face challenges in accurately determining objectness and bounding-box regression due to independent learning processes, leading to suboptimal inference accuracy.
Innovation Solution
An information processing apparatus that integrates an image acquisition unit, GT data acquisition, inference, score calculation, determination, and parameter update units to enhance objectness learning by calculating loss values based on overlap scores and updating neural network parameters using Momentum Stochastic Gradient Descent.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If objectness learning and bounding-box regression learning are performed independently, then the learning process is simpler and faster, but the inference accuracy is reduced
Solution Approach 1:
The patent merges objectness learning and bounding-box regression learning into a unified learning process. The loss function combines both objectives: L = L_objectness + λL_regression, where L_objectness is the binary cross-entropy loss for objectness classification and L_regression is the smooth L1 loss for bounding-box regression. This integrated approach allows both tasks to be optimized simultaneously, improving inference accuracy while maintaining learning efficiency through a single end-to-end training process.
2Ease of manufacture
If anchor boxes are set at the center of corresponding anchors, then the learning process is more straightforward, but the objectness learning does not always align with accurate bounding-box regression
Solution Approach 1:
The patent implements feedback by using the bounding-box regression results to refine objectness learning. The unified loss function allows the regression output to provide feedback to the objectness classification through the combined optimization process. Specifically, the gradient from the regression loss L_regression flows back to update the shared features, which in turn improves objectness prediction. This feedback mechanism ensures that objectness learning and bounding-box regression learning are aligned, improving reliability without complicating the anchor box placement.
Data Source
AI summary
An information processing apparatus configured to learn objectness of an object in an image includes an image acquisition unit configured to acquire an image, a GT data acquisition unit configured to acquire GT data including at least an object area where an object present in the acquired image is present, an inference unit configured to infer a candidate area for the object in the image and a first score indicating objectness for the candidate area based on a learning model, a determination unit configured to determine a second score indicating objectness for the object area based on the inferred candidate area and an area included in the acquired GT data, and an update unit configured to update a parameter for the learning model based on a loss value calculated based on the second score and the inferred first score indicating the objectness for the candidate area.