Source-linked AI summary
Cross-view Asymmetric Metric Learning for Unsupervised Person Re-identification
Hong-Xing Yu, Ancong Wu, Wei-Shi Zheng
TL;DR
Unsupervised RE-ID needs alternatives to expensive cross-view labels and must address camera-specific interference. CAMEL learns view-specific projections with asymmetric metric clustering, and experiments report stronger general performance, especially on large-scale unlabeled datasets.
Problem
Supervised RE-ID requires expensive labeled data across disjoint camera views, while unsupervised methods struggle to model view-specific illumination and occlusion differences.
Method
CAMEL jointly learns an asymmetric metric, view-specific projections, and cluster separations to map cross-view data into a shared space with reduced bias.
Results
CAMEL outperforms existing methods generally, with especially strong performance on large-scale unlabeled RE-ID datasets.
Takeaways & Limitations
Clustering-based asymmetric metric learning can support unsupervised RE-ID while explicitly addressing camera-view differences.
Abstract
from arXiv · showhide
While metric learning is important for Person re-identification (RE-ID), a significant problem in visual surveillance for cross-view pedestrian matching, existing metric models for RE-ID are mostly based on supervised learning that requires quantities of labeled samples in all pairs of camera views for training. However, this limits their scalabilities to realistic applications, in which a large amount of data over multiple disjoint camera views is available but not labelled. To overcome the problem, we propose unsupervised asymmetric metric learning for unsupervised RE-ID. Our model aims to learn an asymmetric metric, i.e., specific projection for each view, based on asymmetric clustering on cross-view person images. Our model finds a shared space where view-specific bias is alleviated and thus better matching performance can be achieved. Extensive experiments have been conducted on a baseline and five large-scale RE-ID datasets to demonstrate the effectiveness of the proposed model. Through the comparison, we show that our model works much more suitable for unsupervised RE-ID compared to classical unsupervised metric learning models. We also compare with existing unsupervised RE-ID methods, and our model outperforms them with notable margins. Specifically, we report the results on large-scale unlabelled RE-ID dataset, which is important but unfortunately less concerned in literatures.
1. Introduction
Unsupervised person re-identification is hindered by expensive cross-view labeling and view-specific interference. CAMEL jointly learns view-specific projections and clustering to create a shared space with reduced bias.
- Person re-identification matches and ranks pedestrians across non-overlapping camera views, but remains challenging because of large intra-class variation and high inter-class similarity.
- Manual cross-view labeling is costly, unreliable at scale, and restrictive for applying supervised methods to large unlabeled surveillance datasets.
- Existing unsupervised methods often treat camera views identically, overlooking illumination and occlusion differences that introduce view-specific interference.
- CAMEL jointly learns an asymmetric metric and cluster separations, using view-specific projections to characterize cross-view images in a shared space.
- The model aims to alleviate view-specific bias in the shared space and thereby improve cross-view matching performance.
- Over 230,000 samples were assembled into a larger unlabeled dataset, where experiments indicate improved scalability for CAMEL relative to existing unsupervised methods.
2. Related Work
Prior RE-ID research is dominated by supervised methods requiring substantial pairwise cross-view labels, while unsupervised approaches generally overlook view-specific bias and clustering.
- Most existing RE-ID models learn metrics, subspaces, view-invariant features, or deep representations using substantial labeled training data.
- Existing unsupervised RE-ID methods treat transformations or quantization across distinct camera views similarly instead of explicitly modeling view-specific bias.
- Earlier unsupervised RE-ID methods rarely use clustering, whereas CAMEL introduces asymmetric metric clustering to characterize data in the learned space.
- Clustering-based metric learning had not previously been attempted in RE-ID, partly because view-specific interference can cause same-view rather than cross-view identity grouping.
3. Methodology
CAMEL learns view-specific projections that map cross-view person images into a shared space, explicitly reducing view-specific interference during unsupervised clustering. Its objective combines within-cluster compactness with cross-view consistency and is optimized by alternating projection updates and k-means clustering.
- Problem formulation: The model learns one projection U^p for each camera view, mapping original M-dimensional features into a shared T-dimensional space.The view-specific projection is selected according to each image’s camera view.
- Asymmetric modelling: The clustering objective gathers similar person images by minimizing within-cluster distances to centroids, relaxing the unavailable identity-label supervision.This adopts a k-means-like objective for the unsupervised setting.
- Asymmetric modelling: View-specific projections address distortions from lighting and occlusions that can dominate cross-view clustering when a universal transformation is used.The method explicitly models and alleviates this view-specific interference.
- Asymmetric modelling: Cross-view consistency regularization limits arbitrary differences among view-specific transformations because images of the same people across views are inherently correlated.The regularizer balances discriminative clustering with reduced view-specific bias.
- Empirical illustration: In SYSU visualization, asymmetric metric clustering alleviates view bias and groups samples of one person together more clearly than symmetric clustering.Shapes denote camera views, while colors denote identities; PCA is used for visualization.
- Optimization: The alternating algorithm fixes cluster assignments to optimize projections by eigen-decomposition, then fixes projections to update assignments with k-means.The objective decreases step by step and is guaranteed to converge.
4. Experiments
Experiments evaluate CAMEL across six RE-ID datasets spanning small to very large scales, using unsupervised protocols and comparisons with competing models. CAMEL consistently outperforms the compared unsupervised methods, while asymmetric modeling is important and the method also transfers across feature types.
- Optimization: The optimization alternates between initializing or updating clusters and solving eigen-decomposition problems to construct or update the projection matrix.The procedure iterates while the objective decrement exceeds ϵ and the maximum iteration count is unreached.
- Comparison with unsupervised models: CAMEL outperforms other unsupervised models on all six datasets under both single-shot and multi-shot settings.Table 2 reports rank-1 CMC accuracies and MAP, while the CMC curves show CAMEL ahead at any rank.
- Comparison with unsupervised models: CAMEL achieves noticeable improvements over clustering-based AML and UsNCA because those models do not model view-specific bias.The authors attribute the advantage to asymmetric modeling that alleviates disturbances during clustering.
- Further evaluations: Replacing CAMEL with a symmetric model worsens performance substantially, indicating that asymmetric modeling is important for cross-view matching.The comparison evaluates the symmetric degradation in the role-of-asymmetric-modeling experiment.
- Further evaluations: CAMEL outperforms Dic, ISR, and L2 distance when using PCA-reduced hand-crafted LOMO features, not only deep-learning-based JSTL features.This evaluates adaptation to a different feature representation.
5. Conclusion
CAMEL demonstrates that clustering-based asymmetric metric learning can make metric learning effective for unsupervised person re-identification, particularly on large-scale unlabelled datasets.
- CAMEL learns view-specific projections to address view-specific interference in unsupervised person re-identification.
- The method builds an asymmetric metric from existing clustering on unlabelled re-identification data.
- Extensive experiments show that CAMEL generally outperforms existing methods, especially on large-scale unlabelled re-identification datasets.