Abstract
The massive attention to surveillance video-based analysis makes vehicle re-identification (Re-ID) a critical area of study. Extracting discriminative visual representations is challenging due to low variance among vehicles sharing the same model, brand, type, and color. This paper proposes Multi-Label-Based Similarity Learning (MLSL), an efficient deep-learning-based model that derives robust vehicle representations. The model features two main components: a multi-label-based similarity learner employing a Siamese network on vehicle ID, color, and type, and a regular CNN-based feature learner for vehicle ID. Validated on VeRi-776, VehicleID, and VERI-Wild datasets, the results demonstrate the superiority of MLSL over multiple state-of-the-art methods.
Introduction
Vehicle Re-ID involves identifying a specific vehicle across non-overlapping surveillance cameras. It is notoriously difficult because of "low-variance" between different identities: two different cars of the same make, model, and color can look virtually identical to standard algorithms.
Existing methods typically fall into two categories: feature learning (treating Re-ID as a massive classification problem) or metric learning (learning a distance metric to pull similar images together and push dissimilar ones apart). However, many existing models are computationally expensive or fail to leverage auxiliary attributes like color and vehicle type that are naturally used by humans to distinguish between similar-looking cars.
Proposed Methodology
The authors propose a dual-branch architecture trained in two stages, utilizing MobileNet v1 as the backbone. Its depthwise separable convolutions make the model lightweight and suitable for real-time applications compared to heavier ResNet or VGG alternatives.
The Two-Part Learner
- Vehicle Feature Learner: A standard branch that uses Softmax classifiers to categorize vehicles by their unique IDs.
- Multi-Label Similarity Learner (MLSL): A Siamese branch that takes pairs of images and computes the element-wise absolute difference between their feature vectors.
Unlike traditional Siamese networks that use a single binary label (same/different ID), MLSL uses three: ID, Color, and Type. This forces the network to learn discriminative features for each attribute. The model is trained using a combination of classification loss and a multi-label contrastive loss, ensuring features are both good for categorization and robust for distance-based matching.
Experimental Evaluation
The model was evaluated on three large-scale datasets:
- VeRi-776: ~50,000 images from 20 cameras; the gold standard for attribute-based Re-ID.
- VehicleID: Over 221,000 images focusing on front and rear viewpoints.
- VERI-Wild: Diverse dataset captured by 174 cameras across various weather and lighting conditions.
Key Results
The MLSL model outperformed several complex state-of-the-art methods, including those using LSTMs and GANs:
- VeRi-776: Achieved a Mean Average Precision (mAP) of 61.13% and a Top-1 accuracy of 90.04%.
- VehicleID: Reached a Top-1 accuracy of 74.21% on the Test-800 subset.
- VERI-Wild: Reached an mAP of 46.32% on the small test subset, a significant improvement over previous benchmarks like FDA-Net (35.11%).
- Ablation Study: Removing the multi-label component (using only ID) dropped the mAP on VeRi-776 from 55.49% to 51.74%, proving the value of the color and type attributes.
Discussion and Insights
- Attribute Synergy: Supervising similarity through color and type acts as a powerful regularizer. Even if two cars have different IDs, if they share a color, the model learns they are "partially similar," refining the feature space more effectively than a simple binary signal.
- Vector vs. Scalar Distance: Keeping the distance between feature vectors as a vector (element-wise difference) before passing it to a sigmoidal mapping layer performed significantly better than collapsing it into a single scalar value too early.
- Efficiency: By using MobileNet v1, the authors demonstrate that top-tier results do not require parameter-heavy backbones if the learning strategy is sufficiently robust.
Conclusion
While efficient and accurate, the current training relies on an online batch sampler to create image pairs, which can be sensitive to implementation. Future work includes applying this multi-label similarity approach to person re-identification—where attributes like clothing, gender, and luggage could serve as auxiliary labels—and exploring the impact of different backbone architectures.
Summary updated