TL;DR
This paper introduces AdsNorm, a novel approach for person re-identification (Re-ID) that aims to generalize well to unseen environments without requiring any target domain data during training. By adaptively leveraging domain-specific normalization statistics and a meta-learning strategy, AdsNorm effectively integrates information from multiple source domains to create a robust model for unseen target domains.
What Problem Does This Paper Address?
Person re-identification (Re-ID) is a crucial task in computer vision, aiming to identify individuals across different camera views. While deep learning has significantly advanced Re-ID performance, existing methods often struggle when deployed in new environments (domains) that differ from their training data. This "domain shift" leads to a sharp drop in accuracy. Traditional solutions like unsupervised domain adaptation (UDA) require unlabeled target data and model updates, which are often impractical or unavailable in real-world scenarios. Domain generalization (DG) aims to train a single model that works well across unseen domains without any target data. However, existing DG methods for Re-ID often focus on learning domain-invariant features, potentially discarding valuable domain-specific information. This paper addresses the challenge of achieving high generalization in person Re-ID by proposing a method that explicitly leverages and adaptively combines domain-specific characteristics from multiple source domains.
Methods
The core of the proposed method, AdsNorm, lies in its "adaptive domain-specific normalization" approach. Instead of learning domain-invariant features, AdsNorm explicitly captures domain-specific characteristics using Batch Normalization (BN) layers. During training, each source domain has its own BN layer statistics (mean and variance) that are updated via a moving average to approximate population statistics.
The system comprises a lightweight ensemble of multiple domain-specific networks, each trained on a different source domain, and a query network. All networks share parameters except for their BN/Instance Normalization (IN) layers. At test time, an image from an unseen target domain is processed by the query network using IN layers, which act as a single-instance BN.
To determine how to best combine information from source domains for a given test image, AdsNorm calculates a "domain relevance score." This score is a combination of two metrics:
- Statistics Distance ($D_{nor}$): Measures the distance between the IN statistics of the test image and the BN statistics of each source domain across all normalization layers, using the Wasserstein distance.
- Feature Affinity ($D_{fet}$): Measures the cosine similarity between the feature extracted by the query network (from the test image) and the domain-specific feature extracted by the corresponding source domain network.
These relevance scores are then used to adaptively weight the features from each source domain network, creating a generalized aggregated feature for the unseen target domain.
To train this ensemble model effectively without access to target domain data, a meta-learning strategy is employed. Source domains are randomly split into meta-train and meta-test sets in each episode. The meta-train phase optimizes the domain-specific networks using standard identification and triplet losses. The meta-test phase then uses the meta-test domain to learn a target-specific feature and an aggregated feature. A novel "relation loss" is introduced to pull the aggregated feature closer to the target-specific feature, encouraging better generalization.
Datasets
The paper evaluates AdsNorm on two protocols for Domain Generalization Person Re-ID:
- Protocol 1: Uses multiple large-scale Re-ID datasets as source domains (Market-1501, DukeMTMC, CUHK02, CUHK03, CUHK-SYSU) and small-scale datasets as unseen target domains (VIPER, PRID, GRID, iLIDS).
- Protocol 2: Employs a leave-one-out setting on four large-scale Re-ID datasets (Market-1501, DukeMTMC, CUHK03, MSMT17), where three are used as source domains and the remaining one as the target domain.
Key Insights and Ideas
- Embracing Domain Specificity: The paper challenges the common DG paradigm of solely pursuing domain-invariant features. It argues that domain-specific characteristics, when adaptively combined, can significantly boost generalization.
- Normalization Statistics as Domain Signatures: Using the statistics of BN/IN layers as a proxy for domain-specific information is an elegant way to capture subtle domain differences without complex architectural changes.
- Meta-Learning for Unseen Domains: The meta-learning approach effectively simulates the unseen target domain scenario during training, allowing the model to learn how to best combine source domain information based on learned relevance.
- Complementary Relevance Metrics: Combining statistical distance ($D_{nor}$) and feature affinity ($D_{fet}$) provides a more robust measure of domain relevance than either metric alone.
Key Results
AdsNorm consistently outperforms state-of-the-art methods across both protocols.
- Protocol 1: AdsNorm achieves the highest R1 accuracy and mAP on all target datasets (VIPER, PRID, GRID, iLIDS), significantly outperforming supervised, UDA, and existing DG methods. For example, it improves R1 accuracy over the second-best DG method DIR-ReID by 6.5% on VIPER.
- Protocol 2: AdsNorm also demonstrates superior performance on large-scale datasets, outperforming methods like M³L by significant margins (e.g., 9.8% R1 accuracy on Market-1501).
- Ablation Studies: The experiments confirm the effectiveness of each component: the ensemble model, the meta-learning strategy, and the combined domain relevance metrics ($D_{nor}$ and $D_{fet}$) all contribute to the overall performance gains.
Limitations and Future Work
- Computational Cost: While described as a "lightweight ensemble," the ensemble nature of the model might still incur higher computational costs during inference compared to a single-network model.
- Meta-Learning Complexity: The meta-learning training strategy, while effective, adds complexity to the training pipeline.
- Future Work: The authors suggest that their approach of adaptively integrating domain-specific features can be extended to other domain generalization tasks beyond person Re-ID. Further exploration into more sophisticated ways to measure domain relevance and combine features could also be beneficial.
Summary updated