Object Boundary Detection Using Extreme Point Area Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods face challenges in achieving both speed and accuracy, particularly in deep learning-based techniques like ExtremeNet and YOLO, due to high computational costs associated with pixel-wise likelihood calculations and insufficient position detection of object extreme points.
Innovation Solution
A method and device utilizing a trained convolutional neural network for parallel execution of extreme point estimation and area estimation, associating feature points on object boundaries, and correcting object areas based on these points, reducing the need for pixel-wise calculations and enhancing detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel-wise likelihood calculation is performed for each pixel at image resolution (ExtremeNet method), then object boundary detection accuracy is improved, but processing time increases significantly
Solution Approach 1:
The image is divided into multiple grids, and only candidate grids containing objects are processed further. This segmentation approach avoids pixel-wise calculation across the entire image, reducing computational load while maintaining boundary detection accuracy for objects in selected grids.
Solution Approach 2:
Instead of performing full pixel-wise likelihood calculation on all pixels, the method performs calculations only on pixels within selected candidate grids that have object presence probability above a threshold. This partial action reduces processing time while maintaining sufficient detection accuracy.
2Productivity
If object area estimation is performed without extreme point calculation (YOLO method), then processing speed is improved, but object position detection accuracy deteriorates
Solution Approach 1:
The method performs preliminary object area estimation using grid-based probability maps first, then selectively calculates extreme points only for candidate objects. This preliminary action enables fast initial detection while preserving the option to refine position accuracy through extreme point calculation when needed.
Solution Approach 2:
Extreme point calculation is applied locally only to candidate grids with high object probability, rather than uniformly across the entire image. This local quality approach maintains high position detection accuracy for detected objects while avoiding unnecessary calculations in background areas, thus preserving processing speed.
Data Source
AI summary
A detection method of detecting a defined object from an image, includes estimating, on the image, an extreme point area including a boundary feature point that satisfies a criterion related to a boundary of the object.