Dynamic Anchor Refinement for Faster Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems, such as Faster R-CNN, face limitations in performance improvement due to the empirical definition of anchors and alignment functions, leading to inefficiencies in object detection and classification.
Innovation Solution
A method and system that improves the region proposal network (RPN) by generating and refining anchors through a series of convolutions and scaling/shifting processes, assigning objectness scores, and using a proposal classifier to enhance accuracy and speed of object detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple anchors are empirically defined for each point in the image, then the coverage of different object scales and aspect ratios is improved, but the number of anchor boxes increases significantly, reducing detection speed
Solution Approach 1:
The patent transforms the static, empirically defined anchor boxes into dynamic anchors that are adaptively generated during the detection process. Instead of pre-defining multiple anchors with fixed scales and aspect ratios, the system generates anchors on-the-fly based on the actual image content and object characteristics, allowing the anchor configuration to adapt dynamically to different detection scenarios without increasing the overall computational burden
Solution Approach 2:
The patent changes the parameters of anchors from fixed, pre-defined values to variable parameters that are learned and adjusted during training. The scale and aspect ratio of anchors are no longer static empirical values but become learnable parameters that adapt to the specific detection task, enabling versatile coverage while maintaining efficient processing speed
2Ease of manufacture
If empirically defined alignment functions are used to match anchors with objects, then the simplicity of implementation is maintained, but the detection accuracy is limited
Solution Approach 1:
The patent replaces the mechanical, hand-crafted alignment functions with a learned neural network-based alignment mechanism. Instead of using fixed geometric transformations defined by empirical formulas, the system employs a neural network to learn the optimal alignment transformations between anchors and objects, achieving superior detection accuracy while maintaining implementation simplicity through end-to-end training
3Productivity
If the number of anchor boxes is reduced to one per point, then the detection speed is improved, but the ability to detect objects with various scales and aspect ratios deteriorates
Solution Approach 1:
The patent performs preliminary action by generating a rich set of candidate anchors during the proposal stage, then selectively refines only the most promising candidates through subsequent convolutional operations. This allows the system to maintain one anchor per point during processing (ensuring speed) while still exploring multiple scale and aspect ratio possibilities through the proposal network's candidate generation
Solution Approach 2:
The patent addresses the scale and aspect ratio diversity problem by introducing a new dimension in the feature space through convolutional operations. Instead of varying anchors in the spatial domain (which would increase their number), the system transforms the single anchor through multiple convolutional layers that operate in the feature space, effectively creating diverse anchor representations without increasing their quantity
Data Source
AI summary
The present invention provides a method comprising the steps of: generating a second anchor on the second convolutional feature map by scaling and shifting a first anchor in the ground-truth box; generating a third convolutional feature map by convolving the second convolutional feature map by means of a second convolution; determining whether the overlap ratio between the ground-truth box and a second single anchor is greater than or equal to a reference value; generating a third anchor by scaling and shifting the second anchor having an overlap ratio that is greater than or equal to the reference value; assigning an objectivity score to the third anchor; and presenting the third anchor, which has an objectivity score that is equal to or greater than a reference value, as a proposal on the third convolutional feature map.