Object Re-Identification Using Self-Dissimilarity Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for object matching in video surveillance struggle with robustly identifying objects across multiple camera views due to variations in photometric properties such as focus, shadows, brightness, contrast, and white balance, leading to inaccurate object recognition and re-identification.
Innovation Solution
The Photometric Invariant Self-dissimilarity Matching (PISM) method determines an appearance signature based on self-dissimilarity between feature maps of image cells, using the signed earth mover's distance metric and soft thresholding to construct an appearance signature that is invariant to photometric changes, enabling accurate object identification across varying conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If appearance-based object matching uses traditional feature vectors (colour, texture, shape) and Bhattacharyya distance, then object re-identification can be performed, but the appearance dissimilarity becomes large when the same object is viewed under different photometric conditions
Solution Approach 1:
The patent transforms the appearance representation from traditional feature vectors to histogram of oriented gradients (HOG) features, and changes the similarity metric from Bhattacharyya distance to self-dissimilarity based on L1 norm. This parameter transformation makes the representation more robust to photometric variations while maintaining discriminative power for object re-identification
Solution Approach 2:
The patent divides the object appearance into multiple local regions and computes HOG features for each region separately. This segmentation allows the method to capture local structural information that is more invariant to global photometric changes, improving robustness while maintaining identification accuracy
2Device complexity
If object matching methods use average intensity and gradients in binary tests, then appearance signatures can be constructed, but the binary differences are sensitive to noise in homogeneous regions and do not characterize the magnitude of difference between pairs of regions
Solution Approach 1:
The patent replaces binary difference tests with continuous L1 norm-based self-dissimilarity measurement. This change preserves the magnitude information of differences between regions, making the appearance signature more robust to noise while maintaining computational efficiency. The self-dissimilarity metric captures both presence and intensity of differences, unlike binary tests that lose magnitude information
3Adaptability or versatility
If self-similarity method is used to characterize geometric shape independently of photometric properties, then photometric invariance is achieved, but the method cannot discriminate different objects with similar shape and cannot match articulated objects under large changes in shape
Solution Approach 1:
The patent applies different treatment to different parts of the object by computing self-dissimilarity for multiple local regions independently. Each region's HOG features capture local structural characteristics that are both photometrically invariant and discriminative. This local quality approach allows the method to distinguish objects with similar global shapes while maintaining photometric invariance
Solution Approach 2:
The patent uses HOG features that can adapt to shape variations through gradient orientation calculations. The oriented gradients naturally handle articulated objects and shape changes by capturing local edge directions, making the appearance signature dynamic and adaptable to large shape variations while maintaining photometric invariance
Data Source
AI summary
A method of identifying an object in an image is disclosed. At least one feature map for each of a plurality of cells in the image is determined. A self-dissimilarity between a first feature map associated with a first one of said cells and a second feature map associated with a second cell, is determined. The self-dissimilarity is determined by determining a sum over thresholds of a difference in area between the first feature map and the second feature map. An appearance signature for the object is formed based on the determined self-dissimilarity. A distance between the appearance signature of the object in the image and appearance signatures of each of a plurality of further objects is determined. The object in the image is identified based on the determined distances.