Person Image Segmentation for Carrying-Object-Independent Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person retrieval systems fail to accurately identify individuals due to the influence of carrying objects, leading to incorrect retrievals or exclusions based on variations in objects carried by the person.
Innovation Solution
A person retrieval system that extracts and processes carrying object images from query images to prevent significant feature extraction, allowing for accurate comparison of person features by calculating and comparing query and registered person feature amounts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If carrying objects are included as part of the person image for retrieval, then the retrieval system can identify persons with their belongings, but the retrieval accuracy deteriorates due to variations in carrying objects affecting feature extraction
Solution Approach 1:
The system segments the person/carrying object image into two distinct parts: the person image and the carrying object image. This segmentation allows the system to extract features from the person independently of the carrying object, thereby maintaining retrieval accuracy while still enabling identification of persons with their belongings. The feature extraction unit extracts person features from the segmented person image, excluding the carrying object features that would otherwise cause variability.
Solution Approach 2:
The system extracts and separates the carrying object image from the combined person/carrying object image. By taking out the carrying object component, the system prevents its features from interfering with person identification. The extracted carrying object image can be processed separately or discarded, while the person image is used for accurate feature extraction and retrieval comparison.
2Adaptability or versatility
If carrying objects are detected as separate parts, then the system can potentially identify objects carried by persons, but the retrieval reliability deteriorates due to incorrect matching when the same person carries different objects
Solution Approach 1:
The system segments the image into person and carrying object components, allowing independent processing. This segmentation ensures that carrying object detection capabilities are maintained while preventing object features from contaminating the person identification process, thereby maintaining reliable retrieval results across different carrying scenarios.
Solution Approach 2:
The carrying object image is extracted and separated from the person image. This extraction prevents the carrying object from being included in the person feature vector, ensuring that the same person will have consistent feature representations regardless of what objects they are carrying, thus maintaining retrieval reliability.
3Ease of manufacture
If the entire person/carrying object image is used for feature extraction, then processing is simpler, but the retrieval precision deteriorates due to significant feature amounts from carrying objects
Solution Approach 1:
The system segments the image into person and carrying object regions, which adds a processing step but enables precise feature extraction from only the relevant person portion. This segmentation approach trades some processing complexity for significant improvement in feature extraction precision by excluding carrying object features.
Solution Approach 2:
The carrying object image is extracted and removed from the feature extraction process. This extraction eliminates the contamination of person features by carrying object features, thereby improving the precision of feature extraction while maintaining reasonable processing complexity through automated image processing techniques.
Data Source
AI summary
A person retrieval system includes a person/carrying object image extraction unit that extracts, from an input image, a person/carrying object image that is an image of a person to be retrieved including an image of a carrying object, a carrying object image extraction unit that extracts an image of the carrying object from the person/carrying object image, a person image calculation unit that calculates a person image that is the person/carrying object image after processing is made on the carrying object part in the person/carrying object image so as to prevent a significant feature amount from being extracted, a person feature amount extraction unit that extracts a person feature amount from the person image, and a determination unit that compares the person feature amount with a registered person feature amount of a registered person, and determines whether the person to be retrieved is the same as the registered person.