Multi-Scale Object Detection With Dual Windows for Lower Compute
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods face challenges in accuracy and computational efficiency due to the variability in object size, particularly when detecting objects in images, leading to increased computational burden and potential recognition errors.
Innovation Solution
A 2-path object recognition structure is employed, dividing the input image into foveated and peripheral regions based on user attention, using dual neural networks to process these regions separately, thereby reducing the size of the image to be processed and improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If different convolution layers are applied to generate multi-scale feature maps based on size variability of target object, then object detection accuracy is improved, but computational burden increases
Solution Approach 1:
The patent divides the image processing into two separate paths: a first path for processing small objects and a second path for processing large objects. Each path uses convolution layers optimized for its specific scale range, avoiding the need to process all scales in both paths, thus reducing overall computational burden while maintaining detection accuracy.
Solution Approach 2:
The patent applies different processing qualities to different regions of the image based on object size. Small objects are processed with higher resolution and more computational resources in the first path, while large objects are processed with lower resolution in the second path, optimizing the balance between accuracy and computational efficiency for each local region.
2Adaptability or versatility
If feature map of each scale is generated irrespective of size of target object, then object detection coverage is improved, but computational complexity increases
Solution Approach 1:
The patent segments the object detection task into two distinct paths based on object size: small objects processed in the first path and large objects processed in the second path. This segmentation allows the system to generate feature maps selectively for each scale rather than generating all scale feature maps universally, reducing computational complexity while maintaining comprehensive detection coverage.
Solution Approach 2:
The patent applies partial action by processing only the necessary scales for each object size category. Instead of generating all possible scale feature maps for all objects, the system generates feature maps at appropriate scales for each object size group, avoiding unnecessary computational operations while ensuring complete detection coverage.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An object detection method includes setting a first window region and a second window region larger than the first window region that correspond to partial regions of different sizes in an input image, downsampling the second window region to generate a resized second window region, detecting a first object candidate from the first window region and a second object candidate from the resized second window region, and detecting an object included in the input image based on one or both of the first object candidate and the second object candidate.