R-CNN Object Detection With Adaptive Pooling for Scale Variation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional R-CNN-based object detectors struggle with precise detection of objects of varying sizes and shapes due to pooling operations being performed with a single scale and aspect ratio, leading to inaccurate object recognition.
Innovation Solution
The proposed method involves a region-based convolutional neural network (R-CNN) that uses multiple anchor boxes with varying scales and aspect ratios to generate ROI candidates, applies pooling operations based on corresponding sizes, and calculates object class and regression losses through backpropagation to enhance detection accuracy without increasing computation or reducing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pooling operations are performed with a single scale and aspect ratio in conventional R-CNN, then the device complexity is reduced, but the measurement precision of object detection deteriorates
Solution Approach 1:
The patent segments the pooling operation into multiple distinct pooling layers, each dedicated to a specific scale and aspect ratio. Instead of using a single pooling operation for all objects, the system divides the detection task into multiple specialized pooling operations that can be applied according to the specific characteristics of detected objects.
Solution Approach 2:
The patent introduces dynamic adaptability by selecting different pooling operations based on the detected object's scale and aspect ratio. The system dynamically chooses the appropriate pooling layer and parameters rather than applying a fixed pooling operation, allowing the detection process to adapt to varying object characteristics.
2Adaptability or versatility
If multiple anchor boxes with varying scales and aspect ratios are used, then the adaptability to different object sizes and shapes is improved, but the device complexity increases
Solution Approach 1:
The patent applies local quality by assigning different pooling operation characteristics to different regions of the feature map based on the local object characteristics. Each pooling layer is optimized for specific scale and aspect ratio ranges, allowing the system to handle diverse object types with specialized local operations rather than a uniform global approach.
Solution Approach 2:
The patent changes the parameters of pooling operations (kernel size, stride, padding) according to the detected object's scale and aspect ratio. By dynamically adjusting pooling parameters based on object characteristics, the system achieves high adaptability without requiring a fixed complex configuration of anchor boxes.
3Measurement precision
If multiple pooling layers with different pooling sizes are applied, then the object detection accuracy is improved, but the computation time increases
Solution Approach 1:
The patent applies partial action by selectively using only the necessary pooling layers based on the detected object's characteristics. Instead of applying all pooling operations to every detected region, the system uses only the relevant pooling layers that match the object's scale and aspect ratio, reducing unnecessary computations while maintaining accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for learning an object detector based on a region-based convolutional neural network (R-CNN) capable of converting modes according to aspect ratios or scales of objects is provided. The aspect ratio and the scale of the objects including traffic lights may be determined according to characteristics, such as distance from the object detector, shapes, and the like, of the object. The method includes steps of: a learning device instructing an RPN to generate ROI candidates; instructing pooling layers to output feature vector; and learn the FC layers and the convolutional layer through backpropagation. In this method, pooling processes may be performed depending on real ratios and real sizes of the objects by using distance information and object information obtained through a radar, a lidar or other sensors. Also, the method can be used for surveillance as humans at a specific location have similar sizes.