Semi-supervised Person Re-identification via Multi-view Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fully-supervised person re-identification methods require extensive manually labeled training data, which is costly and time-consuming, limiting their applicability to new scenarios with large amounts of unlabeled data, while semi-supervised and unsupervised approaches are less competitive.
Innovation Solution
A semi-supervised learning framework that utilizes multiple convolutional neural networks (CNNs) for fine-tuning with labeled data and propagates labels to unlabeled samples through multi-view clustering, enabling accurate label estimation and improving pedestrian retrieval accuracy by integrating CNN model features and using siamese networks for identification and verification losses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fully-supervised learning methods are used for person re-identification, then retrieval accuracy can be improved, but the requirement for large amounts of manually labeled training data increases, making the process expensive and time-consuming
Solution Approach 1:
The system performs preliminary clustering on unlabeled data to generate pseudo-labels before actual training. By pre-organizing unlabeled images into clusters using feature extraction and clustering algorithms, the system creates a structured dataset that can be quickly labeled and used for training, avoiding the need for manual labeling of all training data while maintaining good retrieval accuracy
Solution Approach 2:
The system enables self-service by automatically generating labels for unlabeled data through multi-view clustering. The clustering algorithm automatically assigns labels to unlabeled images based on their feature similarity across multiple views, allowing the system to create its own training dataset without external human intervention, thus reducing the quantity of manually labeled data required
2Quantity of substance
If semi-supervised or unsupervised approaches are used to reduce labeled data requirements, then the cost and time for data labeling decrease, but the retrieval accuracy and performance become less competitive
Solution Approach 1:
The system merges supervised and unsupervised learning approaches by combining a small set of labeled data with a large set of unlabeled data. The multi-view clustering algorithm processes both labeled and unlabeled data together, allowing the system to leverage the structure information from unlabeled data while maintaining the accuracy benefits of supervised learning on labeled data, thus achieving competitive retrieval accuracy with reduced labeling requirements
Solution Approach 2:
The system introduces multi-view clustering as an intermediary mechanism between labeled and unlabeled data. This intermediary process generates pseudo-labels for unlabeled data that are then used to train the retrieval model, bridging the gap between unsupervised pre-processing and supervised fine-tuning, thereby maintaining retrieval accuracy while reducing manual labeling requirements
Data Source
AI summary
A semi-supervised model incorporates deep feature learning and pseudo label estimation into a unified framework. The deep feature learning can include multiple convolutional neural networks (CNNs). The CNNs can be trained on available training datasets, tuned using a small amount of labeled training samples, and stored as the original models. Features are then extracted for unlabeled training samples by utilizing the original models. Multi-view clustering is used to cluster features to generate pseudo labels. Then the original models are tuned by using an updated training set that includes labeled training samples and unlabeled training samples with pseudo labels. Iterations of multi-view clustering and tuning using an updated training set can continue until the updated training set is stable.