Unsupervised Training Sample Selection for Person Re-Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person re-identification models face challenges in adapting to new environments due to the domain shift problem, where the distribution of appearance descriptors in the source and target domains differ significantly, leading to poor performance when deployed in new settings, especially with unsupervised learning methods that lack labeled training samples and are prone to noise and outliers.
Innovation Solution
A method for generating training samples by creating a first representation of unlabelled objects from images, learning a manifold using an unsupervised method, selecting anchor images, determining close and distant candidate images, and generating a training sample from the anchor and match candidate images, which allows for effective updating of the person re-identification model without labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If unsupervised learning methods are used to train person re-identification models without labeled training samples, then the model can be applied to new environments, but the model performance deteriorates due to noise and outliers in the training data
Solution Approach 1:
The system automatically identifies and selects high-quality training samples from unlabelled data by computing similarity metrics between query images and candidate images. The model serves itself by selecting training data without human intervention, using the learned representation to identify suitable samples that will improve its performance in the target domain.
Solution Approach 2:
The system uses a feedback mechanism where the trained model's performance on query images informs the selection of training samples. By computing similarity scores and selecting candidate images that meet specific criteria (close in representation space but distant in image space), the system continuously improves its model through self-generated feedback from its own performance metrics.
2Reliability
If labeled training samples are used to train person re-identification models, then the model performance is improved, but the model cannot adapt to new environments with different domain distributions
Solution Approach 1:
The system performs preliminary actions by first training the model on source domain data, then using this pre-trained model to select training samples from the target domain. This preliminary training provides a foundation that guides the subsequent sample selection process, enabling the model to adapt to new environments while maintaining performance through carefully selected training data.
3Quantity of substance
If all available unlabelled images are used as training samples, then the quantity of training data is maximized, but the quality of training samples decreases due to inclusion of noisy and outlier data
Solution Approach 1:
The system applies local quality by selecting training samples based on their local characteristics in the representation space. Instead of treating all images uniformly, it identifies specific regions (close candidates and distant candidates) with different quality properties and selects samples from these localized regions that meet specific similarity criteria, ensuring high quality training data.
Solution Approach 2:
The system changes parameters by using multiple similarity metrics (cosine similarity, Euclidean distance) and adjusting selection thresholds to identify suitable training samples. By varying these parameters and computing multiple distance metrics, the system can filter out noisy data and select only high-quality training samples that meet the specified criteria.
Data Source
AI summary
A method of generating a training sample for matching unlabelled objects in a sequence of images. A first representation of unlabelled objects is generated from images of a first and second set of images. A second representation of the unlabelled objects is generated using an unsupervised method. An anchor image in the first set is selected. A first set of candidate images in the second set that are close to the anchor image in both the first and second representations, is determined. A second set of candidate images in the second set that are distant from the anchor image in either the first or the second representations, is determined. A match candidate image is selected from the first set or the second set of candidate images. The training sample is generated from at least the anchor image and the match candidate image.