Object Detection Device Using Reliability Thresholds to Reduce Feature Map Reads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using CNNs are bottlenecked by the need for software-based detection processing, which is not accelerated like the convolution operations, leading to inefficiencies due to the need to read feature maps from DRAM.
Innovation Solution
An object detection device and method that acquire metadata including object position and reliability from a CNN, store feature map values, and only read feature map values related to object positions when the reliability exceeds a threshold, thereby reducing unnecessary reads and speeding up detection processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If detection processing is implemented by software to convert feature map values to bounding boxes, then flexibility and ease of implementation are improved, but processing speed deteriorates due to the need to read feature maps from DRAM
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing reliability information (objectness scores) in the feature map before detection processing. This allows the system to quickly identify and skip low-reliability regions during detection, avoiding unnecessary reads from DRAM and accelerating processing while maintaining software-based flexibility
Solution Approach 2:
The patent implements partial action by performing detection processing only on regions where the reliability threshold is met. Instead of processing the entire feature map, the system selectively processes only high-reliability regions, reducing the number of DRAM reads and improving processing speed while maintaining accurate object detection
2Reliability
If all feature map values are read from storage to perform detection processing, then completeness of processing is improved, but processing time increases due to unnecessary reads of low-reliability regions
Solution Approach 1:
The patent extracts and utilizes the reliability information (objectness score) from the feature map to identify and extract only the high-reliability regions for detection processing. By separating high-reliability regions from low-reliability regions, the system avoids reading unnecessary data from storage, reducing processing time while maintaining detection completeness for meaningful objects
Solution Approach 2:
The patent applies partial action by performing detection processing only on regions where the reliability threshold is met. This selective processing approach ensures that all potentially meaningful objects are detected (maintaining completeness) while avoiding wasteful processing of low-reliability regions, thereby reducing overall processing time
Data Source
AI summary
An object detection device 10 including a metadata acquisition unit 103, a storage unit 104, and a feature map value acquisition unit 105 is provided. The metadata acquisition unit 103 acquires metadata including at least a position and reliability of an object included in an image from a convolutional neural network into which the image is input. The storage unit 104 stores a feature map value group which is an output result of the convolutional neural network. The feature map value acquisition unit 105 reads a feature map value related to the position of the corresponding object from the storage unit 104 to obtain the position of the object only when the reliability obtained by reading a feature map value, which is related to the reliability in the feature map value group stored in the storage unit 104, from the storage unit 104 exceeds a predetermined threshold value.