Instance Segmentation via Hierarchical Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current instance segmentation methods in computer vision face challenges in achieving accurate and efficient segmentation, particularly in autonomous driving and domestic robotics, due to high computational complexity and limited use of hierarchical information.
Innovation Solution
The proposed method involves performing feature extraction on images using a neural network to output features at multiple hierarchies, extracting region features from instance candidate regions, and fusing these features to obtain a fusion feature for improved instance segmentation, which reduces complexity and enhances accuracy by leveraging information from different semantic hierarchies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature extraction is performed at multiple hierarchical levels, then segmentation accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent divides the feature extraction process into multiple hierarchical levels, where each level processes features at a different scale. This segmentation allows the system to capture both fine-grained details and global contextual information separately, improving segmentation accuracy while managing computational complexity through structured organization of processing stages.
Solution Approach 2:
The patent introduces a hierarchical dimension to the feature extraction process, organizing features across multiple levels of abstraction. By adding this dimensional structure, the system can efficiently manage computational complexity while accessing information at different scales, resolving the contradiction between accuracy and complexity.
2Measurement precision
If region features are extracted and fused from multiple hierarchies, then instance segmentation accuracy is improved, but processing time increases
Solution Approach 1:
The patent performs preliminary feature extraction and organization at multiple hierarchical levels before the actual segmentation decision is made. By preparing and fusing region features in advance across different hierarchies, the system reduces the computational burden during the final segmentation step, thereby improving accuracy while minimizing additional processing time.
3Loss of information
If hierarchical feature fusion is performed, then use of hierarchical information is improved, but computational requirements increase
Solution Approach 1:
The patent extracts and fuses region features from specific hierarchical levels that are most relevant to instance segmentation. By selectively taking out and processing only the necessary hierarchical features rather than all possible features, the system improves hierarchical information utilization while reducing overall computational requirements.
Data Source
AI summary
An instance segmentation method includes: performing feature extraction on an image via a neural network to output features at at least two different hierarchies; extracting region features corresponding to at least one instance candidate region in the image from the features at the at least two different hierarchies, and fusing region features corresponding to a same instance candidate region, to obtain a first fusion feature of each instance candidate region; and performing instance segmentation based on each first fusion feature, to obtain at least one of an instance segmentation result of the corresponding instance candidate region or an instance segmentation result of the image.