Unsupervised Cross-dataset Person Re-identification by Transfer Learning of Spatial-Temporal Patterns

Abstract: Most of the proposed person re-identification algorithms conduct supervised training and testing on single labeled datasets with small size, so directly deploying these trained models to a large-scale real-world camera network may lead to poor performance due to underfitting. It is challenging to incrementally optimize the models by using the abundant unlabeled data collected from the target domain. To address this challenge, we propose an unsupervised incremental learning algorithm, TFusion, which is aided by the transfer learning of the pedestrians' spatio-temporal patterns in the target domain. Specifically, the algorithm firstly transfers the visual classifier trained from small labeled source dataset to the unlabeled target dataset so as to learn the pedestrians' spatial-temporal patterns. Secondly, a Bayesian fusion model is proposed to combine the learned spatio-temporal patterns with visual features to achieve a significantly improved classifier. Finally, we propose a learning-to-rank based mutual promotion procedure to incrementally optimize the classifiers based on the unlabeled data in the target domain. Comprehensive experiments based on multiple real surveillance datasets are conducted, and the results show that our algorithm gains significant improvement compared with the state-of-art cross-dataset unsupervised person re-identification algorithms.

Unsupervised Cross-dataset Person Re-identification by Transfer Learning of Spatial-Temporal Patterns

The paper presents a novel approach for unsupervised cross-dataset person re-identification (Re-ID), addressing the prominent challenge of deploying trained models from small, labeled datasets to large-scale, unlabeled camera networks, where differences in data distribution can significantly affect the performance. The proposed TFusion algorithm leverages transfer learning combined with spatial-temporal pattern integration to optimize classifier effectiveness incrementally and unsupervisedly, showcasing significant improvements over existing methodologies.

Spatial-Temporal Patterns and Visual Features

The TFusion algorithm incorporates spatial-temporal patterns of pedestrian movements as integral components alongside visual features commonly used in Re-ID systems. By transferring a visual classifier trained on a labeled source dataset to the target domain, the model learns these patterns from the statistical properties of temporal intervals and camera configurations of image pairs judged to contain the same individual. This involves estimating probabilities associated with movements across the camera network without requiring explicit prior information on camera locations or pedestrian movement assumptions.

Bayesian Fusion Model

A key element of the TFusion framework is the Bayesian fusion model, which integrates the learned spatial-temporal patterns with visual features to enhance classification accuracy. This fusion model calculates the matching probability using a Bayesian approach that combines visual similarity scores with spatial-temporal likelihoods. The paper demonstrates that this fusion approach, under specified error conditions, can reduce error rates compared to using visual features alone. This methodology is particularly effective in scenarios where spatial-temporal patterns are distinct and predictable, such as structured environments with defined pedestrian paths.

Incremental Optimization via Learning-to-Rank

TFusion further incorporates a learning-to-rank scheme to refine its model iteratively by predicting and optimizing ranking orders based on the fusion-based similarity scores. By using ranked results from the fusion model, the visual classifier is continuously updated to improve its accuracy, leveraging the unlabeled data in the target environment. This mutual promotion strategy significantly enhances both classifiers' performance through iterative learning cycles.

Experimental Results and Implications

The implementation and testing of TFusion on various benchmark datasets illustrate its superiority compared to state-of-the-art unsupervised transfer algorithms and its competitive performance against supervised counterparts. Notably, in environments like metro stations with pronounced movement patterns, TFusion shows remarkable success, essentially outperforming supervised methods despite relying solely on unlabeled data. This underscores the potential of incorporating spatial-temporal dynamics in large-scale Re-ID applications, suggesting theoretical and practical pathways for advancing algorithmic designs that better adapt to real-world deployment.

Future Developments in AI

The concepts introduced in this work open avenues for further exploration into spatio-temporal models in machine learning, especially in domains reliant on environmental dynamics and pattern recognition. Future development could involve refining the estimation techniques for spatial-temporal data or integrating new sensor modalities to enhance the model's robustness and adaptability. Additionally, such methodologies suggest broader applications beyond surveillance, potentially influencing how AI systems interpret complex, multi-dimensional data across diverse fields.