R-CNN Object Detector Small Object Detection via Pyramidic Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional R-CNN methods are inadequate for detecting small objects in images as they rely on the smallest feature map generated after multiple convolution operations, which is insufficient for precise object representation.
Innovation Solution
The method involves generating multiple feature maps through sequential convolution operations and upsampling blocks to create pyramidic feature maps, allowing for the detection of small objects without additional computational load by using feature maps of varying scales.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple feature maps of varying scales are generated through upsampling blocks, then detection precision for small objects is improved, but device complexity increases
Solution Approach 1:
The feature extraction process is segmented into multiple parallel pathways, each producing feature maps at different scales. Instead of using a single feature map, the system divides the detection task across multiple scaled feature maps, allowing small objects to be detected on appropriately-scaled maps while maintaining overall system precision without excessive complexity
Solution Approach 2:
The patent introduces a scale dimension to the feature maps by applying upsampling operations. Feature maps are generated not only through sequential convolutional layers but also by expanding their spatial dimensions through upsampling blocks, creating a multi-scale representation that improves small object detection while managing complexity through structured dimensional transformation
2Reliability
If multiple feature maps are used for detection, then small object detection capability is improved, but computational load increases
Solution Approach 1:
Different regions of the feature pyramid are utilized based on their scale-appropriate qualities. Small objects are detected on upsampled feature maps where they occupy sufficient pixel density, while larger structures are detected on original-scale maps. This local optimization ensures computational resources are focused where they provide maximum detection value, improving small object detection capability without uniformly increasing computational load across all detection tasks
Solution Approach 2:
The system applies upsampling and feature map generation selectively rather than uniformly to all detection scenarios. By generating additional scaled feature maps only when and where small objects are present in the input image, the system achieves improved small object detection capability while avoiding the computational overhead of processing full multi-scale feature pyramids in every detection case
Data Source
AI summary
A method for learning an object detector based on an R-CNN by using a first to an n-th filter blocks respectively generating a first to an n-th feature maps through convolution operations in sequence, and a k-th to a first upsampling blocks respectively coupled with the first to the n-th filter blocks is provided. The method includes steps of: a learning device instructing the k-th upsampling block to the first upsampling block to generate a (k−1)-st pyramidic feature map to the first pyramidic feature map respectively; instructing an RPN to generate each ROI corresponding to each candidate region, and instructing a pooling layer to generate a feature vector; and learning parameters of the FC layer, the k-th to the first upsampling blocks, and the first to the n-th filter blocks by backpropagating a first loss generated by referring to object class information, object regression information, and their corresponding GTs.