Abstract
Motorcycles are often related to transit and criminal issues due to their abundance in urban traffic. Despite their importance, motorcycles are a seldom addressed problem in the computer vision community. We credit this problem to the lack of large-scale datasets and strong baseline models. Therefore, we present the first large-scale Motorcycles Re-Identification (MoRe) dataset. MoRe consists of 3,827 individuals (i.e., the set of motorbikes and motorcyclists) captured by ten surveillance cameras placed in Brazil's urban traffic scenarios. Furthermore, we evaluate a deep learning model trained using well-known training tricks from the object re-identification literature to present a strong baseline for the motorcycle re-identification (ReID) problem. More importantly, we highlight some crucial problems in this topic, such as the influence of distractors and domain shift. Experimental results demonstrate the effectiveness of the strong baseline model with an increase of at least 19.27 p.p. in the rank-1 when compared to the state-of-the-art in the BPReID dataset. Finally, we present some insights regarding the information learned by the strong baseline model when computing the similarities between motorcycle images.
Introduction
Object Re-Identification (ReID) involves recognizing the same individual across non-overlapping camera views. While significant progress has been made for pedestrians and four-wheel vehicles, motorcycles remain under-addressed. This gap is critical for public safety, as motorcycles are frequently involved in urban crimes and accidents. Unlike cars, motorcycles often lack front license plates and possess varied rider attributes—such as helmets, clothing, and backpacks—making them harder to track with standard systems. Existing datasets are often too small or limited to controlled campus settings, failing to capture the complexity of real-world urban traffic.
The MoRe Dataset
The MoRe (Motorcycle Re-Identification) dataset is the first large-scale effort to address these challenges. It contains 17,619 images of 3,827 distinct identities captured by 10 surveillance cameras in Brazil. The dataset is designed to reflect the messy reality of urban surveillance, featuring low resolution, motion blur, and heavy occlusion.
To simulate real-world conditions, the gallery includes 3,478 "distractors"—images without corresponding probes. This addition significantly impacts performance, dropping Rank-1 accuracy by over 5%, which underscores the importance of evaluating models in realistic scenarios rather than "lab-clean" environments.
Methodology
The researchers established a strong baseline by adapting modern ReID "bag of tricks" to a ResNet50 backbone pre-trained on ImageNet. Key optimizations included:
- Label Smoothing and Warmup Learning Rate: To stabilize training and prevent overconfidence.
- BNNeck: Separating ID classification and metric learning features.
- Center Loss: Reducing intra-class variance to ensure images of the same identity are closer in feature space.
- Metric Learning: Quadruplet Loss was found to outperform Triplet and MSML losses.
Explainable AI (xAI) visualizations confirmed that the model effectively learns features from both the motorcycle (color, model) and the rider (helmet, backpacks), leveraging the unique "pair" nature of motorcycle ReID.
Experimental Results
The proposed baseline achieved a Rank-1 accuracy of 83.41% on the existing BPReID dataset, a 19.27 percentage point improvement over previous handcrafted methods. On the new MoRe dataset, the model achieved 83.41% Rank-1 accuracy and 86.38% mAP (mean Average Precision).
A key finding was the impact of domain shift. Models trained on campus-based datasets (BPReID) performed poorly on urban data (MoRe), highlighting the need for diverse training data. Additionally, performance varied by camera pair; transitions with similar viewpoints reached 85% Rank-1 accuracy, while difficult angles dropped to 40%.
Conclusion
MoRe provides a vital resource for advancing motorcycle ReID. While the baseline shows strong performance, challenges like motion blur and cross-city generalization (domain shift) remain. Future research should explore Unsupervised Domain Adaptation (UDA) and the use of temporal information from video sequences to improve robustness in diverse urban environments.
Summary updated