Graph Sampling Based Deep Metric Learning for Generalizable Person Re-Identification (2104.01546v4)

Abstract: Recent studies show that, both explicit deep feature matching as well as large-scale and diverse training data can significantly improve the generalization of person re-identification. However, the efficiency of learning deep matchers on large-scale data has not yet been adequately studied. Though learning with classification parameters or class memory is a popular way, it incurs large memory and computational costs. In contrast, pairwise deep metric learning within mini batches would be a better choice. However, the most popular random sampling method, the well-known PK sampler, is not informative and efficient for deep metric learning. Though online hard example mining has improved the learning efficiency to some extent, the mining in mini batches after random sampling is still limited. This inspires us to explore the use of hard example mining earlier, in the data sampling stage. To do so, in this paper, we propose an efficient mini-batch sampling method, called graph sampling (GS), for large-scale deep metric learning. The basic idea is to build a nearest neighbor relationship graph for all classes at the beginning of each epoch. Then, each mini batch is composed of a randomly selected class and its nearest neighboring classes so as to provide informative and challenging examples for learning. Together with an adapted competitive baseline, we improve the state of the art in generalizable person re-identification significantly, by 25.1% in Rank-1 on MSMT17 when trained on RandPerson. Besides, the proposed method also outperforms the competitive baseline, by 6.8% in Rank-1 on CUHK03-NP when trained on MSMT17. Meanwhile, the training time is significantly reduced, from 25.4 hours to 2 hours when trained on RandPerson with 8,000 identities. Code is available at https://github.com/ShengcaiLiao/QAConv.

• The paper introduces the Graph Sampling (GS) method that creates informative mini-batches by linking classes to their nearest neighbors.
• It reduces training time and computational cost by eliminating heavy class memory modules and using efficient pairwise learning.
• Cross-dataset results show a 25.1% increase in Rank-1 accuracy, highlighting the method's impact on generalization in person re-identification.

Overview of "Graph Sampling Based Deep Metric Learning for Generalizable Person Re-Identification"

This paper proposes a novel approach named Graph Sampling (GS), aimed at improving deep metric learning in the context of person re-identification. Person re-identification is a significant area in computer vision, which involves matching a person’s image across a number of different camera views. The authors pinpoint the need to enhance generalization capabilities of person re-identification systems, particularly when interfaced with large-scale training data.

Key Contributions and Methodology

1. Introduction of Graph Sampling (GS): The primary contribution of the paper is the GS method for mini-batch sampling in person re-identification tasks. Unlike traditional random sampling strategies like the PK sampler, GS constructs a graph connecting each class to its nearest neighbors at the start of each training epoch. Thus, it ensures that mini-batches contain the most informative and challenging examples — an approach theorized to yield superior discriminative learning.
2. Reduction in Training Time and Computational Cost: The GS technique is particularly effective in reducing the memory and time overhead, making it feasible to train models on large-scale datasets more efficiently. It eliminates the reliance on class memory modules which are resource-intensive, by switching to pairwise learning within mini-batches.
3. Performance Boost in Generalization: The proposed methodology significantly enhances the performance on cross-dataset evaluations. For instance, leveraging GS to train on the Market-1501 dataset resulted in a 25.1% improvement in Rank-1 accuracy on the MSMT17 dataset.
4. Baseline Adaptation and Evaluation: The authors improved an existing robust baseline (QAConv) and demonstrated that the GS-integrated version outperforms the baseline consistently in various cross-dataset scenarios. As demonstrated, the GS method provides more challenging examples which facilitate better learning of discriminative features, thereby enhancing both within-dataset accuracy and cross-dataset generalizability.

Numerical Results

The implementation of GS shows substantial gains in efficiency: training time on the RandPerson dataset decreased from 25.4 hours to just 2 hours when the system was trained on 8,000 identities. On the benchmark datasets, such as MSMT17 and Market-1501, GS leads to compelling advancements in Rank-1 accuracy, underscoring its potential as a robust sampler in deep metric learning frameworks.

Implications and Future Directions

The implications of the GS method extend beyond re-identification tasks. As a general batch sampler, it could potentially be adapted to other domains requiring efficient sample selection from large datasets, such as face recognition or image retrieval. By addressing training efficiency and model generalization, this paper sets the stage for future investigations into scalable metric learning techniques that balance complexity and performance.

The paper also invites questions on how such methodologies can be integrated into unsupervised domain adaptation contexts and whether GS can effectively manage noisy data — an area ripe for further exploration. As the community seeks to design more generalizable machine learning models, the graph sampling strategy put forth by the authors represents an instrumental advance in achieving these objectives.