[Paper Review] Discriminative Feature and Dictionary Learning with Part-aware Model for Vehicle Re-identification

Huibing Wang, Jinjia Peng+3 more|arXiv (Cornell University)|Mar 16, 2020

Video Surveillance and Tracking Methods34 references4 citations

TL;DR

This paper proposes a Triplet Center Loss-based Part-aware Model (TCPM) that enhances vehicle re-identification by learning discriminative local features through spatial partitioning of vehicle images into horizontal and vertical parts. By integrating external memory modules to model global feature distributions and using triplet-center loss to enforce intra-class compactness and inter-class separability, TCPM achieves state-of-the-art performance on VehicleID and VeRi-776, with mAP gains of up to 1.29% and Rank-1 improvements of 1.12%.

ABSTRACT

With the development of smart cities, urban surveillance video analysis will play a further significant role in intelligent transportation systems. Identifying the same target vehicle in large datasets from non-overlapping cameras should be highlighted, which has grown into a hot topic in promoting intelligent transportation systems. However, vehicle re-identification (re-ID) technology is a challenging task since vehicles of the same design or manufacturer show similar appearance. To fill these gaps, we tackle this challenge by proposing Triplet Center Loss based Part-aware Model (TCPM) that leverages the discriminative features in part details of vehicles to refine the accuracy of vehicle re-identification. TCPM base on part discovery is that partitions the vehicle from horizontal and vertical directions to strengthen the details of the vehicle and reinforce the internal consistency of the parts. In addition, to eliminate intra-class differences in local regions of the vehicle, we propose external memory modules to emphasize the consistency of each part to learn the discriminating features, which forms a global dictionary over all categories in dataset. In TCPM, triplet-center loss is introduced to ensure each part of vehicle features extracted has intra-class consistency and inter-class separability. Experimental results show that our proposed TCPM has an enormous preference over the existing state-of-the-art methods on benchmark datasets VehicleID and VeRi-776.

Motivation & Objective

• Address the challenge of distinguishing vehicles with similar appearances due to shared designs or manufacturers in surveillance systems.
• Improve vehicle re-identification performance by focusing on fine-grained local part features rather than relying solely on global appearance features.
• Enhance intra-class consistency and inter-class separability of local vehicle features through a novel loss function and memory-based feature modeling.
• Develop a scalable framework that leverages spatial partitioning and global dictionary learning to boost discriminative feature learning in vehicle re-identification.

Proposed method

• Partitions vehicle feature maps into multiple horizontal and vertical regions to extract localized discriminative features.
• Introduces an external memory module that stores and maintains the global central feature vector of each part across all identities, forming a global dictionary.
• Employs a triplet-center loss function that uses the global central feature vector as an anchor to optimize feature embeddings for intra-class compactness and inter-class margin.
• Uses a two-branch network architecture: one branch for horizontal partitioning and another for vertical partitioning, enabling multi-directional feature learning.
• Applies the memory module to stabilize feature learning by modeling the distribution of local features across the entire dataset.
• Combines global and local feature learning by fusing representations from multiple part regions into a unified embedding for re-identification.

Experimental results

Research questions

• RQ1 Can spatial partitioning of vehicle features into horizontal and vertical parts improve discriminative feature learning for vehicle re-identification?
• RQ2 Does incorporating an external memory module that models global feature distributions enhance intra-class consistency and re-ID accuracy?
• RQ3 To what extent does triplet-center loss improve inter-class separability and intra-class compactness compared to standard triplet loss?
• RQ4 How does the proposed part-aware model compare to global-feature-only baselines and existing state-of-the-art methods on standard benchmarks?
• RQ5 Can the model generalize across varying test set sizes and challenging conditions such as viewpoint and illumination variations?

Key findings

• The proposed TCPM model achieves a mAP of 85.13% and Rank-1 accuracy of 81.96% on the VeRi-776 dataset, outperforming the baseline by 5.06 percentage points in mAP.
• On the VehicleID dataset, TCPM achieves a mAP of 74.59% and Rank-1 of 93.98%, representing a 1.29% improvement in mAP over the baseline.
• The addition of triplet-center loss (OneBranch+Mem+TC) leads to a 1.05% mAP gain and 0.89% Rank-1 gain on VeRi-776 compared to OneBranch+Mem.
• The two-branch architecture (TwoBranch+Mem+TC) improves mAP by 0.51%, 0.59%, and 0.79% on VehicleID test sets of size 800, 1600, and 2400, respectively.
• Visualization results confirm the model’s robustness to viewpoint and illumination variations, with high accuracy in retrieving correct matches across diverse conditions.
• The external memory module effectively stabilizes feature learning and improves generalization by modeling the global distribution of local part features.