Multi-Domain Learning and Identity Mining for Vehicle Re-Identification (2004.10547v2)

Abstract: This paper introduces our solution for the Track2 in AI City Challenge 2020 (AICITY20). The Track2 is a vehicle re-identification (ReID) task with both the real-world data and synthetic data. Our solution is based on a strong baseline with bag of tricks (BoT-BS) proposed in person ReID. At first, we propose a multi-domain learning method to joint the real-world and synthetic data to train the model. Then, we propose the Identity Mining method to automatically generate pseudo labels for a part of the testing data, which is better than the k-means clustering. The tracklet-level re-ranking strategy with weighted features is also used to post-process the results. Finally, with multiple-model ensemble, our method achieves 0.7322 in the mAP score which yields third place in the competition. The codes are available at https://github.com/heshuting555/AICITY2020_DMT_VehicleReID.

• The paper presents a novel multi-domain learning approach combined with Identity Mining to improve vehicle re-identification performance.
• It adapts the BoT-BS baseline with a BNNeck design, achieving 95.8% rank-1 accuracy and 79.9% mAP on the Veri-776 benchmark.
• The work employs a tracklet-level re-ranking method that boosts performance, securing 0.7322 mAP and a top-3 finish in the AI City Challenge 2020.

Multi-Domain Learning and Identity Mining for Vehicle Re-Identification

The paper in question presents a comprehensive solution to the Vehicle Re-Identification (ReID) task as part of the AI City Challenge 2020. This research primarily targets improving vehicle ReID within a city-scale multi-camera environment using both real-world and synthetic datasets. The authors present their novel approach, which comprises a multi-domain learning strategy combined with an innovative Identity Mining method to bolster performance beyond conventional baselines.

Vehicle ReID is a critical area within intelligent transportation systems (ITS), tasked with identifying vehicles across multiple camera feeds without relying on identifiable information like license plates. The methodology adopted for this paper draws on recent advancements in deep learning, notably from the person ReID domain, to address challenges posed by the significant domain gaps between synthetic and real-world data.

Key Contributions and Methodologies

1. Baseline Model Development: The authors employ a robust baseline, dubbed BoT-BS, originally developed for person ReID, and adapt it to vehicle ReID applications. This baseline integrates Bag of Tricks and includes mechanisms such as the BNNeck, which reconcile possible inconsistencies between ID loss functions and triplet loss during training. Such an approach has shown substantial efficacy, evidenced by 95.8% rank-1 and 79.9% mAP on the Veri-776 benchmark.
2. Multi-Domain Learning (MDL): To address the domain disparity between training datasets (real and synthetic), the paper introduces MDL. Here, models are initially pre-trained on a combined dataset of real-world vehicles and a select subset of synthetic vehicle data, followed by fine-tuning specifically on the real-world data. This strategy ensures that low-level features common in both domains are leveraged to improve model performance across disparate data sources.
3. Identity Mining (IM): The proposed Identity Mining method provides a more impactful way to annotate test data with pseudo labels for unsupervised learning, as opposed to traditional clustering methods like k-means. Through local optimization, IM generates highly accurate pseudo labels by selectively identifying and clustering samples based on learned feature distances.
4. Tracklet-Level Re-Ranking: Utilizing tracklet information available for each instance, the authors apply a tracklet-level re-ranking strategy with weighted features. This approach comprehensively evaluates multiple frames of a vehicle across a track sequence, improving robustness and accuracy beyond standard image-to-image re-ranking methods.

Numerical Results and Leaderboard Performance

The paper reports substantial quantitative improvements, achieving 0.7322 mAP on the competition's leaderboard, securing third place among the 41 submissions. This was accomplished through model ensemble strategies, coupling five distinct models, each trained with variations of the proposed methodology. The re-ranking strategy alone significantly enhanced mAP by approximately 5%.

Implications and Future Directions

This work underscores the challenging nature of vehicle ReID tasks involving empirical and synthetic data. The methodologies introduced by the authors—multi-domain learning and Identity Mining—highlight future potential in improving model generalization across varied domains. These strategies could become vital components of intelligent transportation systems as urban environments grow increasingly reliant on automated surveillance technologies.

Future work could focus on improving the global optimization aspect of the Identity Mining strategy to achieve better pseudo-label accuracy and extend these techniques to other domains where domain shift is prevalent. Moreover, investigating automated ways to tune model hyperparameters and adaptively select synthetic data subsets may yield further performance gains in large-scale ReID tasks.