Source-linked AI summary
Learning a Discriminative Null Space for Person Re-identification
Li Zhang, Tao Xiang, Shaogang Gong
TL;DR
Person re-identification must match people across changing camera views using high-dimensional features despite having few matched training samples, creating the small sample size problem. The paper learns a discriminative null space that collapses each identity’s images while preserving class separation, using a closed-form, parameter-free approach. Across five benchmarks and both supervised settings, it reports state-of-the-art performance, often by a large margin.
Problem
High-dimensional re-identification features are trained from relatively few matched samples, making within-class scatter matrices singular and forcing dimensionality reduction or regularisation.
Method
The paper learns a discriminative null space in which same-identity images collapse to one point while between-class variance remains non-zero.
Results
The method achieves state-of-the-art performance on five benchmarks under both fully supervised and semi-supervised settings.
Takeaways & Limitations
A closed-form NFST model provides a simple, efficient, parameter-free way to address the SSS problem in person re-identification.
Takeaways & Limitations
The reported MCML comparison used different features, and the authors could not obtain a meaningful solution with higher-dimensional LOMO features.
Abstract
from arXiv · showhide
Most existing person re-identification (re-id) methods focus on learning the optimal distance metrics across camera views. Typically a person's appearance is represented using features of thousands of dimensions, whilst only hundreds of training samples are available due to the difficulties in collecting matched training images. With the number of training samples much smaller than the feature dimension, the existing methods thus face the classic small sample size (SSS) problem and have to resort to dimensionality reduction techniques and/or matrix regularisation, which lead to loss of discriminative power. In this work, we propose to overcome the SSS problem in re-id distance metric learning by matching people in a discriminative null space of the training data. In this null space, images of the same person are collapsed into a single point thus minimising the within-class scatter to the extreme and maximising the relative between-class separation simultaneously. Importantly, it has a fixed dimension, a closed-form solution and is very efficient to compute. Extensive experiments carried out on five person re-identification benchmarks including VIPeR, PRID2011, CUHK01, CUHK03 and Market1501 show that such a simple approach beats the state-of-the-art alternatives, often by a big margin.
1. Introduction
Person re-identification is difficult because appearance changes across cameras and similar clothing creates imposters, while high-dimensional features and few matched samples produce the SSS problem. The paper addresses this by learning a discriminative null space that collapses same-identity images and reports strong benchmark performance.
- Cross-camera changes in view angle, pose, illumination, and background clutter make person re-identification largely unsolved.
- Thousands-dimensional appearance features are typically learned from only hundreds of matched training samples, creating the small sample size problem.Matched-pair collection is laborious, so the sample size can be much smaller than the feature dimension.
- The proposed discriminative null space collapses same-class training images to one point while retaining non-zero between-class variance.This simultaneously drives within-class scatter to zero and maximizes the Fisher discriminative criterion.
- The null-space approach has a closed-form solution, fixed dimension, no tuning parameters, no preprocessing dimension reduction, and efficient computation.
- Five-benchmark experiments show the approach beats state-of-the-art methods, often by a large margin.
2. Related Work
Prior re-identification work develops increasingly high-dimensional features and learns metrics or discriminative subspaces, but these approaches share the SSS challenge. This paper revisits NFST for re-id and extends it to use unlabelled data in semi-supervised learning.
- Re-identification features have grown to thousands or tens of thousands of dimensions, yet remain vulnerable to cross-view appearance changes.Examples cited in the paper use 5,138- and 26,960-dimensional representations.
- Metric learning and discriminative subspace learning are closely related because projected Euclidean distance equals a positive-semidefinite Mahalanobis distance.Both therefore encounter the SSS problem, often requiring PCA-based dimension reduction or regularisation.
- NFST addresses SSS by restricting projections to directions with zero within-class distance and positive between-class distance.Its single-point-per-class structure limits extensions such as graph-Laplacian modelling.
- The paper uses conventional supervised NFST for its closed-form, parameter-free solution and extends it to semi-supervised learning with unlabelled data.
3. Methodology
The method learns a discriminative null space that collapses each identity’s training samples while preserving between-identity separation, avoiding singular within-class scatter in the SSS regime. It has fixed dimensionality, a closed-form solution, efficient computation, and extensions for kernelisation and semi-supervised learning.
- Foley-Sammon Transform: FST learns projection directions that maximise the Fisher discriminant criterion using between-class and within-class scatter matrices.When Sw is non-singular, it yields a C − 1 dimensional discriminative subspace through a generalised eigen-problem.
- Problem: Person re-id represents samples with very high-dimensional appearance features, while matched training data are scarce, making d > N and Sw singular.FST consequently encounters numerical problems in the small sample size setting.
- Null Foley-Sammon Transform: NFST collapses training samples from each class to one point while keeping the resulting class points distinct, producing zero within-class and positive between-class scatter.The resulting directions are called Null Projecting Directions.
- Learning the Discriminative Null Space: The discriminative null space contains precisely C − 1 Null Projecting Directions, so its dimension is fixed at m = C − 1.This construction guarantees the best training-data separability under the Fisher criterion.
- Learning the Discriminative Null Space: Learning the null space reduces to an eigen-problem with a closed-form solution, efficient computation, and no free parameter to tune.The method computes directions in the shared space between the total-scatter and within-class-scatter null spaces.
- Kernelisation and Semi-supervised Learning: NFST can be kernelised to address nonlinear person appearance, and test images are matched using Euclidean distances after projection into the learned space.A semi-supervised extension uses unlabelled data to construct cross-view correspondences and iteratively create pseudo-classes.
4. Experiments
Experiments across five re-identification benchmarks evaluate fully supervised, semi-supervised, and runtime performance. The null-space method generally outperforms competing approaches, remains effective under severe SSS conditions, and has practical test-time cost.
- Experimental setup: The experiments cover VIPeR, PRID2011, CUHK01, CUHK03, and Market1501, using CMC accuracy and, for Market1501, mAP.VIPeR and PRID2011 also support semi-supervised evaluation with one third of training data labelled.
- Fully supervised results: 42.28% Rank 1 accuracy on VIPeR with a single feature type exceeds XQDA’s 40.00%.With score-level fusion, the method beats the nearest rival by over 5% on Rank 1.
- Fully supervised results: On CUHK01 and CUHK03, the method beats compared methods at low ranks or under detector-generated boxes, while deep learning becomes more competitive on the larger CUHK03 dataset.Fusing two feature types further boosts performance, and detector boxes reduce matching accuracy for all compared methods.
- Fully supervised results: On Market1501, the method significantly outperforms alternatives with the same LOMO features under both single-query and multi-query settings and both evaluation metrics.This occurs despite 12,936 training samples making SSS least severe among the evaluated datasets.
- Semi-supervised results: 24.70% Rank 1 accuracy remains after reducing the labelled training data to one third, versus 29.80% in the fully supervised setting on PRID2011.The authors attribute the more graceful decrease partly to self-training exploiting unlabelled data.
- Running cost: Training is most efficient among the compared methods, while testing is slower than XQDA but faster than kLFDA and MFA.The authors report that test time over 3,368 queries is more than adequate for real-time applications.
5. Conclusion
The paper proposes NFST-based discriminative null-space learning to address the SSS problem in person re-identification, achieving state-of-the-art performance across five benchmarks in supervised and semi-supervised settings.
- NFST learns a discriminative null space to address the small sample size problem in person re-identification.The model is simpler than existing re-id models, has a closed-form solution, and requires no parameters to tune.
- Extensive experiments on five benchmarks show state-of-the-art performance under both fully supervised and semi-supervised settings.