Object Detector Learning Using Target Object Estimating and Merging Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CNN-based object detectors face challenges in efficiently detecting objects of various sizes in an image without significantly increasing computational load.
Innovation Solution
The method involves using a target object estimating network and a target object merging network to learn the parameters of an object detector. This is achieved by generating feature maps, object proposals, and detection information through convolutional layers, RPNs, pooling layers, and FC layers, while merging proposals and detection information across different scales to optimize object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the CNN processes the entire image at high resolution to detect small objects, then the detection precision for small objects is improved, but the computational load and processing time increase significantly
Solution Approach 1:
The image is divided into multiple regions of interest (ROIs) based on object proposals from the RPN. Instead of processing the entire high-resolution image, the CNN only processes these segmented ROIs at different scales. This segmentation approach maintains detection precision for small objects while significantly reducing the overall computational load and processing time.
Solution Approach 2:
The patent introduces a multi-scale processing dimension by generating and processing images at different scales (original scale, enlarged scale for small objects). The target object estimating network determines which scales to process based on object proposals, adding a scale dimension to the processing pipeline. This allows small objects to be detected with high precision while avoiding the need to process the entire image at maximum resolution.
2Adaptability or versatility
If multiple images at different scales are generated and processed to detect objects of various sizes, then the adaptability to different object sizes is improved, but the device complexity increases
Solution Approach 1:
The RPN (Region Proposal Network) performs preliminary action by generating object proposals and estimating target object locations before the main CNN processing. The target object estimating network uses these proposals to determine which ROIs and scales need further processing. This preliminary action simplifies the overall system complexity by pre-identifying regions of interest, avoiding the need for complex multi-scale processing of the entire image.
Solution Approach 2:
The system dynamically adjusts the processing scale and regions based on the target object estimating network's analysis of object proposals. Instead of a fixed multi-scale processing pipeline, the system adapts its processing dynamics to focus computational resources only on relevant regions and scales, reducing structural complexity while maintaining adaptability to different object sizes.
3Reliability
If the RPN generates numerous object proposals to ensure all objects are captured, then the detection coverage is improved, but the quantity of data to be processed increases
Solution Approach 1:
The target object estimating network extracts and selects only the most relevant object proposals from the RPN's output based on estimated target object locations. Instead of processing all proposals generated by the RPN, this extraction process filters down to a manageable subset of high-priority ROIs, maintaining comprehensive detection coverage while significantly reducing the quantity of data that needs to be processed by the CNN.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for learning parameters of an object detector based on a CNN adaptable to customer's requirements such as KPI by using a target object estimating network and a target object merging network is provided. The CNN can be redesigned when scales of objects change as a focal length or a resolution changes depending on the KPI. The method includes steps of: a learning device instructing convolutional layers to generate a k-th feature map by applying convolution operations to a k-th manipulated image which corresponds to the (k-1)-th target region on an image; and instructing the target object merging network to merge a first to an n-th object detection information, outputted from an FC layer, and backpropagating losses generated by referring to merged object detection information and its corresponding GT. The method can be useful for multi-camera, SVM(surround view monitor), and the like, as accuracy of 2D bounding boxes improves.