[Paper Review] AttKGCN: Attribute Knowledge Graph Convolutional Network for Person Re-identification

Bo Jiang, Xixi Wang+1 more|arXiv (Cornell University)|Nov 24, 2019

Video Surveillance and Tracking Methods44 references4 citations

TL;DR

This paper proposes AttKGCN, a novel end-to-end framework that integrates attribute knowledge graph convolutional networks with visual feature learning for person re-identification. By modeling attribute co-occurrence dependencies via a directed attribute knowledge graph and leveraging graph convolutional networks for inter-dependent attribute classification, AttKGCN achieves state-of-the-art performance on benchmark datasets, significantly improving both attribute prediction and Re-ID accuracy.

ABSTRACT

Discriminative feature representation of person image is important for person re-identification (Re-ID) task. Recently, attributes have been demonstrated beneficially in guiding for learning more discriminative feature representations for Re-ID. As attributes normally co-occur in person images, it is desirable to model the attribute dependencies to improve the attribute prediction and thus Re-ID results. In this paper, we propose to model these attribute dependencies via a novel attribute knowledge graph (AttKG), and propose a novel Attribute Knowledge Graph Convolutional Network (AttKGCN) to solve Re-ID problem. AttKGCN integrates both attribute prediction and Re-ID learning together in a unified end-to-end framework which can boost their performances, respectively. AttKGCN first builds a directed attribute KG whose nodes denote attributes and edges encode the co-occurrence relationships of different attributes. Then, AttKGCN learns a set of inter-dependent attribute classifiers which are combined with person visual descriptors for attribute prediction. Finally, AttKGCN integrates attribute description and deeply visual representation together to construct a more discriminative feature representation for Re-ID task. Extensive experiments on several benchmark datasets demonstrate the effectiveness of AttKGCN on attribute prediction and Re-ID tasks.

Motivation & Objective

• To address the challenge of learning discriminative person re-identification features under large appearance variations such as pose, illumination, and occlusion.
• To exploit the co-occurrence relationships among person attributes, which are often ignored in existing methods, to improve attribute prediction and Re-ID performance.
• To unify attribute learning and visual representation learning into a single end-to-end framework that jointly optimizes both tasks.
• To develop a novel attribute knowledge graph (AttKG) that encodes co-occurrence patterns between attributes as directed edges between nodes.
• To demonstrate that incorporating structured attribute dependencies via graph convolution improves both attribute recognition and person re-identification accuracy.

Proposed method

• Constructs a directed attribute knowledge graph (AttKG) where nodes represent attributes (e.g., 'Female', 'Bag') and edges encode learned co-occurrence probabilities between attribute pairs.
• Employs a graph convolutional network (GCN) module to learn inter-dependent attribute classifiers by aggregating neighborhood information from the AttKG.
• Integrates the output of the GCN-based attribute classifiers with deep visual features from a pre-trained CNN backbone to generate a unified, discriminative person representation.
• Applies an attribute re-weighting module that dynamically adjusts the contribution of each attribute to the final representation based on its confidence and relevance.
• Trains the entire network end-to-end using a multi-task loss combining Re-ID classification loss and attribute prediction cross-entropy loss.
• Uses a learnable parameter λ to balance the contributions of Re-ID and attribute prediction losses during joint optimization.

Experimental results

Research questions

• RQ1 Can modeling attribute co-occurrence dependencies through a knowledge graph improve attribute prediction performance in person re-identification?
• RQ2 Does integrating a graph-based attribute learning module into a deep Re-ID model lead to more discriminative visual representations?
• RQ3 How does the proposed AttKGCN framework compare to state-of-the-art attribute-guided Re-ID methods in terms of mAP and Rank-1 accuracy?
• RQ4 What is the impact of the attribute re-weighting module on the final Re-ID performance and attribute prediction reliability?
• RQ5 How sensitive is the AttKGCN model to hyperparameters such as the loss balance weight λ and the number of GCN layers?

Key findings

• AttKGCN achieves 94.4% Rank-1 and 85.5% mAP on the Market-1501 dataset, outperforming the baseline model without attribute learning.
• On DukeMTMC-reID, AttKGCN achieves 87.8% Rank-1 and 77.4% mAP, demonstrating consistent gains across multiple benchmarks.
• The model shows robustness to hyperparameter λ, with stable performance in the range of 4–6, and achieves optimal results at λ=5.
• The ablation study confirms that both the GCN-based attribute module and the re-weighting mechanism are essential, as removing either leads to significant performance drops.
• The attribute prediction accuracy of AttKGCN surpasses competing methods like APR and AANet, especially in average accuracy across all attributes.
• Visualization results show that AttKGCN produces more clearly separated feature embeddings per identity compared to the baseline, indicating stronger discriminative capacity.