TL;DR
This paper introduces EMA-GR, a novel framework for gait recognition that explicitly handles "Out-of-Gallery" (OOG) queries – cases where a person's gait isn't present in the reference database. By leveraging Evidential Deep Learning and Dempster-Shafer Theory, EMA-GR quantifies recognition uncertainty, allowing it to accurately identify OOG cases and significantly improve robustness compared to existing methods.
What Problem Does This Paper Address?
Traditional gait recognition systems assume that every query (probe) sample will have a corresponding identity within the reference database (gallery). This assumption breaks down in real-world scenarios where a person might be observed but not be in the system's database. When faced with such "Out-of-Gallery" (OOG) probes, existing methods often incorrectly assign an identity from the gallery, leading to erroneous results. The paper addresses this critical limitation by proposing a unified framework that can not only identify individuals present in the gallery (In-Gallery or IG queries) but also reliably detect and flag OOG queries.
Methods
The core of the EMA-GR framework lies in its uncertainty-aware approach. Instead of just outputting a match score, EMA-GR quantifies the confidence in that match. This is achieved through an "evidence collector," typically a small neural network, that processes the features of a probe sequence and its closest match from the gallery. This evidence is then used to parameterize a Dirichlet distribution, drawing from the Dempster-Shafer Theory of Evidence (DST). This distribution allows the model to output three scores: a matching score, a non-matching score, and an uncertainty score.
The framework is designed to be agnostic to the underlying gait recognition backbone. This means it can be plugged into existing state-of-the-art gait recognition models without requiring them to be retrained from scratch. For training the evidence collector, the authors employ a "hard mining" strategy to construct challenging positive (same identity) and negative (different identity) pairs, ensuring the model learns to handle ambiguous cases.
Datasets
The proposed EMA-GR framework was evaluated on several standard gait recognition datasets:
- OUMVLP: A large-scale dataset used for evaluating performance under various OOG percentages.
- CASIA-B: A popular dataset used for gait recognition, also evaluated with varying OOG rates.
- Gait3D: A large-scale, in-the-wild dataset.
Additionally, to demonstrate the generalizability of EMA-GR, experiments were conducted on datasets for other identity retrieval tasks:
- VERI-Wild: A large-scale vehicle re-identification dataset.
- VC-Clothes: A synthetic person re-identification dataset.
Key Insights and Ideas
- The critical need for OOG detection: The paper highlights that ignoring OOG cases leads to significant performance degradation and incorrect identifications in practical applications.
- Uncertainty as a key indicator: By quantifying the uncertainty in the match status, EMA-GR can effectively distinguish between a true match and an OOG query. This is a departure from traditional methods that only focus on similarity scores.
- Agnosticism is powerful: The ability to integrate EMA-GR with existing gait recognition backbones without retraining is a major advantage, making it easily adoptable and improving the robustness of current systems.
- Dempster-Shafer Theory for robust uncertainty: Applying DST and Dirichlet distributions provides a principled way to model and quantify uncertainty, allowing the model to express "I don't know" when necessary.
- Hard mining for challenging cases: Training the evidence collector with carefully constructed hard positive and negative pairs is crucial for learning to handle ambiguous and OOG scenarios effectively.
Key Results
- EMA-GR significantly improves the performance of existing gait recognition methods on datasets with OOG queries.
- On the OUMVLP dataset, EMA-GR achieved a 51.26% improvement when the OOG query rate was around 50%, surpassing state-of-the-art methods by a substantial margin.
- The framework consistently outperforms baseline methods across various OOG percentages on OUMVLP and CASIA-B datasets.
- Experiments on VERI-Wild and VC-Clothes demonstrate EMA-GR's generalizability to other identity retrieval tasks, showing significant improvements in OOG scenarios.
- The proposed method maintains high recognition accuracy for In-Gallery queries while effectively identifying Out-of-Gallery probes.
Limitations and Future Work
- Performance degradation in IG scenarios: The paper notes that EMA-GR can sometimes lead to a slight decrease in accuracy for In-Gallery (IG) queries, especially when combined with backbones that produce less discriminative features. This is attributed to the framework's preference for rejecting high-risk predictions to avoid OOG misclassifications.
- Dataset size impact: The authors speculate that training on smaller datasets like CASIA-B might limit the model's ability to learn sufficiently discriminative features, potentially impacting performance in high OOG percentage scenarios.
- Future Work: The paper suggests extending the EMA-GR framework to other tasks, which has already been demonstrated with vehicle and person re-identification. Further research could explore more advanced evidence collection mechanisms or investigate how to further minimize the performance drop in IG scenarios.
Summary updated