Source-linked AI summary
RAM: A Region-Aware Deep Model for Vehicle Re-Identification
Xiaobin Liu, Shiliang Zhang, Qingming Huang, Wen Gao
TL;DR
Vehicle Re-ID remains difficult when different vehicles share similar global appearances, motivating features that capture distinctive local details. RAM combines global, regional, and attribute branches with joint supervision, and achieves the best reported performance on VeRi and VehicleID against the compared methods.
Problem
Vehicles with the same maker and model can look similar globally, while existing global descriptions may lose discriminative local details needed for vehicle Re-ID.
Method
RAM jointly learns global and three-overlapping-region features, while using vehicle identity, model, maker, and color cues through multiple classification tasks.
Results
RAM achieves the best reported performance on both VeRi and VehicleID compared with the other evaluated methods.
Takeaways & Limitations
Combining regional and attribute cues produces more discriminative vehicle features and supports distinguishing visually similar vehicles.
Abstract
from arXiv · showhide
Previous works on vehicle Re-ID mainly focus on extracting global features and learning distance metrics. Because some vehicles commonly share same model and maker, it is hard to distinguish them based on their global appearances. Compared with the global appearance, local regions such as decorations and inspection stickers attached to the windshield, may be more distinctive for vehicle Re-ID. To embed the detailed visual cues in those local regions, we propose a Region-Aware deep Model (RAM). Specifically, in addition to extracting global features, RAM also extracts features from a series of local regions. As each local region conveys more distinctive visual cues, RAM encourages the deep model to learn discriminative features. We also introduce a novel learning algorithm to jointly use vehicle IDs, types/models, and colors to train the RAM. This strategy fuses more cues for training and results in more discriminative global and regional features. We evaluate our methods on two large-scale vehicle Re-ID datasets, i.e., VeRi and VehicleID. Experimental results show our methods achieve promising performance in comparison with recent works.
1. INTRODUCTION
Vehicle Re-ID must distinguish individual vehicles even when vehicles sharing a maker and model look globally similar. RAM addresses this gap by combining global, regional, and attribute cues, with experiments on VeRi and VehicleID showing promising performance.
- Problem: Vehicle Re-ID identifies reappearing vehicle instances in camera networks, unlike related tasks that classify fine-grained vehicle categories.The task supports applications including surveillance analysis, vehicle tracking, and intelligent transportation.
- Motivation: Vehicles sharing the same maker and model can have similar global appearances, while their distinguishing differences often occur in local regions.Previous global-image approaches may lose discriminative local details, motivating explicit regional cue extraction.
- Approach: RAM jointly learns deep features from the whole vehicle image and three overlapping local regions through a multi-branch architecture.Its branches include Conv, BN, Regional, and Attribute branches, whose features are concatenated for vehicle instance identification.
- Evaluation: Experiments evaluate RAM on the VeRi and VehicleID large-scale vehicle Re-ID datasets.The paper reports promising performance in comparison with recent works.
- Approach: Color and model cues are jointly used to train RAM alongside vehicle identity supervision.The resulting concatenated feature is reported to achieve promising performance compared with recent methods.
2. PROPOSED METHOD
RAM uses shared convolutional layers followed by complementary global, regional, and attribute branches. Each branch receives task-specific supervision, and the resulting features are concatenated for vehicle Re-ID.
- Network Structure: Five shared convolutional layers generate feature maps that feed Conv, BN, Attribute, and Regional branches.The branches produce global features fc and fb, attribute feature fa, and regional features frt, frm, and frb.
- Global Branches: The Conv branch learns a global identity feature from pooled whole-image feature maps, while the BN branch adds contextual global cues.Batch normalization depresses highly activated regions and increases the visibility of other regions, producing complementary features.
- Regional Branch: The Regional branch divides feature maps into three overlapping top, middle, and bottom regions and learns a feature from each part.Training each region to identify vehicles encourages extraction of discriminative local details.
- Attribute Branch: Attribute features encode model, maker, and color cues that can complement visual features and be more robust to viewpoint, illumination, and background changes.The Attribute branch learns these cues through an attribute classification task.
- Training: RAM trains each branch with an individual softmax classification task and combines the losses in a weighted objective.The Region loss contains three equally weighted classification losses for the different regions.
- Training: The model is trained step by step, adding BN, Region, and Attribute branches after first training the Conv branch.Shared convolutional layers are then fine-tuned across multiple classification tasks, and the final features represent vehicles from different aspects.
3. EXPERIMENTS
RAM was evaluated on VeRi and VehicleID through branch-wise experiments and comparisons with recent methods. Adding complementary global, regional, and attribute cues improved performance, with RAM achieving the best reported comparisons on both datasets.
- Datasets and evaluation: Experiments evaluated RAM on the VeRi and VehicleID vehicle Re-ID datasets using mAP, Top-1, and Top-5 metrics.VeRi uses mAP, Top-1, and Top-5; VehicleID uses Top-1 and Top-5.
- Branch-wise evaluation: Each ablation step added one branch, progressing from the Conv baseline through BN, BN+R, and the final RAM model.The concatenated feature was tested after each branch addition.
- VeRi ablation: On VeRi, concatenating global features [fc; fb] outperformed either feature alone, indicating that the BN branch supplied complementary information.The BN model’s fc also outperformed the baseline fc.
- VeRi ablation: 0.609 mAP was achieved by [fc; fb; fr], exceeding 0.601, 0.590, and 0.593 for single-region alternatives.The result supports jointly fusing all three regional features rather than relying on one region.
- VeRi ablation: The final feature [fc; fb; fr; fa] improved VeRi mAP by 6.5% over the baseline feature.It was the best-performing feature among those compared in Table 1.
- VehicleID evaluation: On VehicleID, adding more branches consistently improved performance, and RAM achieved the best Rank-1 and Rank-5 accuracies among the tested concatenations.Retrieved examples also showed RAM outperforming the baseline global feature for visually similar vehicles and viewpoint variations.
- Comparison with state of the art: RAM achieved the best performance on both datasets in comparisons with recent methods, outperforming SCPL by 3.2% mAP on VeRi and OIF by 13.5% mAP on VeRi and 0.7% Top-1 on VehicleID.The comparison includes methods with deeper networks, additional spatio-temporal cues, more training data, or more complex structures.
4. CONCLUSION
RAM combines global, regional, and attribute features for vehicle Re-ID, and experiments on two large-scale vehicle datasets demonstrate promising performance.
- RAM extracts regional features from three overlapped local regions in addition to global features.
- The Region branch encourages attention to local details, producing more discriminative features for vehicle Re-ID.
- RAM jointly trains an Attribute branch to generate features that may be more robust to viewpoint variations.
- Experiments on two large-scale vehicle datasets demonstrate RAM’s discriminative features and promising performance.