Person Re-identification with Deep Similarity-Guided Graph Neural Network (1807.09975v1)

Abstract: The person re-identification task requires to robustly estimate visual similarities between person images. However, existing person re-identification models mostly estimate the similarities of different image pairs of probe and gallery images independently while ignores the relationship information between different probe-gallery pairs. As a result, the similarity estimation of some hard samples might not be accurate. In this paper, we propose a novel deep learning framework, named Similarity-Guided Graph Neural Network (SGGNN) to overcome such limitations. Given a probe image and several gallery images, SGGNN creates a graph to represent the pairwise relationships between probe-gallery pairs (nodes) and utilizes such relationships to update the probe-gallery relation features in an end-to-end manner. Accurate similarity estimation can be achieved by using such updated probe-gallery relation features for prediction. The input features for nodes on the graph are the relation features of different probe-gallery image pairs. The probe-gallery relation feature updating is then performed by the messages passing in SGGNN, which takes other nodes' information into account for similarity estimation. Different from conventional GNN approaches, SGGNN learns the edge weights with rich labels of gallery instance pairs directly, which provides relation fusion more precise information. The effectiveness of our proposed method is validated on three public person re-identification datasets.

• The paper presents a novel SGGNN approach that integrates inter-gallery image relations to improve similarity estimation in person Re-ID.
• It employs a complete graph where each node is a probe-gallery pair, using learned gallery similarities to guide effective message passing.
• Robust performance is demonstrated with top-1 accuracies of up to 95.3% on CUHK03, outperforming conventional methods.

Deep Similarity-Guided Graph Neural Network for Person Re-identification

The paper presents a novel approach to the person re-identification (Re-ID) problem using a Deep Similarity-Guided Graph Neural Network (SGGNN). The method addresses a fundamental challenge in Re-ID: improving the accuracy of similarity estimation between probe and gallery image pairs by considering the relation information that exists among all pairs within a dataset.

Traditional Re-ID models typically evaluate the similarity of probe-gallery image pairs independently. This approach often overlooks the potential for leveraging inter-gallery image relations, which could enhance the robustness of similarity estimations, particularly for hard samples. The proposed method, SGGNN, mitigates this limitation by forming a graph that integrates information between different image pairs and leverages these relationships for improved similarity estimation.

Methodological Insights

The SGGNN is a deep learning framework that formulates person Re-ID as a node-focused graph application. It constructs an undirected complete graph where each node represents a probe-gallery pair. The nodes are initially assigned pairwise relation features derived from a simple Siamese CNN model. The primary innovation lies in the edge formulation, where edges capture the similarity relations among gallery images. The similarity relations, directly supervised with identity labels, guide the message-passing mechanism within the network.

A key contribution of SGGNN is its refinement of probe-gallery similarity estimations through deep messages propagated across the graph. The message-passing mechanism updates node relation features by aggregating information from other nodes, weighted by learned gallery-gallery similarities. This stands in contrast to conventional Graph Neural Networks (GNNs), which often use compatibility scores that are indirectly learned. This direct exploitation of gallery image labels for weighting the message passing is a significant advancement, enabling more precise feature fusion.

Numerical Performance

The efficacy of the proposed SGGNN is validated on three large-scale datasets: CUHK03, Market-1501, and DukeMTMC. SGGNN demonstrates superior performance in terms of mean Average Precision (mAP) and cumulative matching characteristics (CMC), significantly outperforming state-of-the-art methods. Results are compelling, with top-1 accuracies reaching up to 95.3% on CUHK03, 92.3% on Market-1501, and 81.1% on DukeMTMC.

Implications and Future Directions

Practically, SGGNN's ability to incorporate inter-gallery similarities in the training phase presents a paradigm shift for Re-ID systems, shifting from simple pairwise measures to a more interconnected perspective. This can dramatically reduce error rates in surveillance and other applications where distinguishing between similar individuals is critical.

Theoretically, the integration of graph neural network architectures embedded with similarity guidance could spur advancements not only in Re-ID but also in other computer vision domains facing similar pairwise assessment challenges. As the design of graph neural networks evolves, future research may explore further refinements, such as dynamic graph construction based on incoming data or integration with temporal sequence analysis for video-based Re-ID tasks.

In summary, the SGGNN framework is a robust and meticulously crafted approach that advances the Re-ID task, leveraging graph-based learning to encapsulate holistic relational information among images. Its exceptional numerical results across multiple datasets affirm its effectiveness and potential for broader application.