Object Detection Visualization with Dynamic Neural Network Layer Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models, particularly 1-stage and 2-stage type models, face challenges in identifying the most suitable layers for visualization of detection results using GradCAM, as the optimal layer for visualization varies based on the target object and its size.
Innovation Solution
A detection result analysis device that calculates an evaluation value for each layer of the object detection model using a heat map representing the activeness degree per pixel within the detection region, and selects suitable layers based on these values to generate a synthesis map for accurate visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the pooling layer after narrowing down the RoI is used for visualization in a 2-stage type model, then the visualization is suitable for GradCAM, but the layer suitability varies depending on the type and size of the target object in a 1-stage type model
Solution Approach 1:
The patent dynamically determines the suitable layer for visualization by calculating evaluation values based on heat maps and detection regions. Instead of using a fixed pooling layer, the system adjusts the layer selection parameter according to the specific detection results, object type, and size, thereby achieving both visualization accuracy and adaptability across different objects
Solution Approach 2:
The system automatically identifies the most suitable layer for visualization by evaluating heat map characteristics and detection region properties. The evaluation value calculation unit self-determines the optimal layer without requiring manual configuration or pre-defined layer selection rules, enabling the system to adapt to different objects autonomously
2Measurement precision
If the pooling layer after narrowing down the RoI is used for visualization, then the visualization works for 2-stage type models, but it is not always best suitable even for 2-stage type models under different conditions
Solution Approach 1:
The patent incorporates feedback mechanisms by calculating evaluation values based on heat maps and detection regions, then using these evaluations to select the most suitable layer. This feedback loop ensures that the layer selection is continuously optimized based on the actual detection results, improving both visualization accuracy and reliability across different conditions
Solution Approach 2:
The system transitions from a static layer selection approach to a dynamic one where the suitable layer is determined in real-time based on the specific detection results. The evaluation value calculation unit dynamically adjusts the layer selection according to the heat map characteristics and detection region properties, ensuring reliable performance across varying conditions
Data Source
AI summary
An evaluation value calculation unit (22) focus on, as a target layer, each of a plurality of layers in an object detection model which detects a target object included in image data and which is constituted using a neural network, and calculates an evaluation value of the target layer from a heat map representing an activeness degree per pixel in the image data obtained from an output result of the target layer, and from a detection region where the target object is detected. A layer selection unit (23) selects at least some layers out of the plurality of layers on a basis of the evaluation value.