Source-linked AI summary
Intra-Inter Camera Similarity for Unsupervised Person Re-Identification
Shiyu Xuan, Shiliang Zhang
TL;DR
Unsupervised person Re-ID needs reliable pseudo-labels despite camera-related distribution discrepancies. The paper decomposes similarity computation into intra-camera CNN-feature matching followed by inter-camera classification-score matching, achieving 89.5% rank-1 accuracy on Market1501 and outperforming recent unsupervised methods by 9+%.
Problem
Camera distribution discrepancies make cross-camera feature similarities unreliable for pseudo-label generation in unsupervised person Re-ID.
Method
The method generates intra-camera pseudo-labels from CNN features, then uses multi-branch classification scores as a more robust inter-camera similarity representation.
Results
89.5% rank-1 accuracy is reported on Market1501 without annotations, outperforming recent unsupervised works by 9+%.
Takeaways & Limitations
Progressive intra-camera and inter-camera pseudo-label optimization produces strong unsupervised person Re-ID performance across multiple datasets.
Takeaways & Limitations
The method uses pre-defined clustering numbers in both stages, making the clustering number a critical parameter for pseudo-label generation.
Abstract
from arXiv · showhide
Most of unsupervised person Re-Identification (Re-ID) works produce pseudo-labels by measuring the feature similarity without considering the distribution discrepancy among cameras, leading to degraded accuracy in label computation across cameras. This paper targets to address this challenge by studying a novel intra-inter camera similarity for pseudo-label generation. We decompose the sample similarity computation into two stage, i.e., the intra-camera and inter-camera computations, respectively. The intra-camera computation directly leverages the CNN features for similarity computation within each camera. Pseudo-labels generated on different cameras train the re-id model in a multi-branch network. The second stage considers the classification scores of each sample on different cameras as a new feature vector. This new feature effectively alleviates the distribution discrepancy among cameras and generates more reliable pseudo-labels. We hence train our re-id model in two stages with intra-camera and inter-camera pseudo-labels, respectively. This simple intra-inter camera similarity produces surprisingly good performance on multiple datasets, e.g., achieves rank-1 accuracy of 89.5% on the Market1501 dataset, outperforming the recent unsupervised works by 9+%, and is comparable with the latest transfer learning works that leverage extra annotations.
1. Introduction
The paper addresses unreliable cross-camera pseudo-labels in unsupervised person Re-ID by decomposing similarity computation into intra-camera and inter-camera stages. Its complete approach reports strong performance without annotations, including 89.5% rank-1 accuracy on Market1501.
- Motivation: Cross-camera appearance variation creates camera-domain gaps that make same-identity samples difficult to match reliably.These gaps also hinder learning high intra-class feature similarity.
- Method: The method first computes CNN-feature similarity within each camera to generate intra-camera pseudo-labels.These pseudo-labels train a multi-branch CNN whose branches correspond to camera-specific classifiers.
- Method: It then uses classification scores from the first stage as feature vectors for more robust inter-camera similarity computation.The score representation is intended to alleviate camera-domain discrepancies during cross-camera clustering.
- Evaluation: The approach is evaluated on multiple person Re-ID datasets using only images and camera labels from each training set.The reported evaluation uses CMC and mAP metrics.
- Results: 89.5% rank-1 accuracy is achieved on Market1501 without annotations, exceeding recent unsupervised methods by 9+%.The authors also report performance above several transfer-learning methods using extra annotations.
- Contribution: The two-stage decomposition progressively seeks better pseudo-labels and combines complementary intra-camera and inter-camera training difficulties.The paper presents this as a more reasonable alternative to directly predicting pseudo-labels across cameras.
2. Related Work
Related work in unsupervised person Re-ID includes distribution alignment, GAN-based image translation, and pseudo-label generation. The paper is most closely related to pseudo-label methods but differs by progressively seeking reliable cross-camera similarity rather than relying on simpler matching strategies.
- Unsupervised Person Re-ID: Unsupervised person Re-ID methods commonly use distribution alignment, GAN-based style transfer, or pseudo-label generation.These categories address unsupervised learning through feature alignment, image translation, or similarity-based labeling.
- Pseudo-Label Methods: Pseudo-label methods generate labels from sample-similarity rules and then train the Re-ID model with those labels.Clustering and KNN search are representative ways to assign similar labels to similar images.
- Domain Adaptation and Generalization: Domain adaptation and generalization research uses normalization strategies such as BN, IN, and IBN-Net to improve CNN generalization across domains.These methods target feature or model generalization to multiple or unseen domains.
- Positioning: The paper differs from related work by progressively eliminating negative influences of pose variation and camera-related discrepancies when seeking reliable similarity.A closely related method uses extra within-camera identity labels and classification results to find cross-camera candidates.
3. Methodology
The method progressively generates pseudo-labels through complementary intra-camera and inter-camera similarity computations. It first trains shared features with camera-specific clustering, then uses classification-score similarity to improve cross-camera clustering and feature robustness.
- Overall framework: The framework alternates intra-camera and inter-camera training to optimize the ReID feature with complementary losses.Intra-camera clustering uses CNN features within each camera; inter-camera clustering uses classification scores across all samples.
- Problem formulation: The method models person appearance as identity-related factors affected by camera settings and stochastic conditions such as pose, illumination, and occlusion.These factors complicate identifying images of the same person across cameras.
- Intra-camera training: Intra-camera training clusters each camera subset using CNN-feature similarity, assigns cluster IDs as pseudo-labels, and trains a shared backbone with camera-specific classifier branches.Sharing the backbone across camera branches enhances feature robustness to camera-specific environments.
- Inter-camera training: Inter-camera training concatenates classification scores from multiple classifiers and uses their Jaccard similarity to estimate whether samples from different cameras share an identity.The resulting inter-camera similarity guides clustering and subsequent optimization with the inter-camera loss.
- Adaptive normalization: The method also proposes AIBN by linearly fusing instance-normalization and batch-normalization statistics.AIBN is introduced to combine the advantages of IN and BN during CNN training.
- Similarity refinement: The inter-camera stage produces the most reliable similarity after intra-camera and inter-camera training progressively improve feature discriminability.The initial cross-camera similarity overlaps with different-identity similarities, making clustering unreliable.
4. Experiments
Experiments on three person Re-ID datasets evaluate the proposed two-stage similarity framework, its components, and hyperparameters. The complete method performs strongly, while ablations support intra-camera similarity, inter-camera similarity design, and AIBN-based generalization.
- Experimental Setup: The method is evaluated on DukeMTMC-ReID, Market1501, and MSMT17 using only training images and camera labels, with CMC and mAP evaluation.The implementation uses a ResNet-50 backbone, staged training, clustering, and dataset-specific training settings.
- Ablation Study: 71.6% and 62.9% rank-1 accuracy are achieved on Market1501 and DukeMTMC-ReID, respectively, with only intra-camera training.Adding inter-camera training without addressing camera distribution gaps can decrease DukeMTMC-ReID performance.
- Ablation Study: AIBN improves performance on both evaluated datasets and increases direct-transfer rank-1 accuracy by 5.6% on Market1501 and 11.3% on DukeMTMC-ReID.The reported results attribute these gains to improved generalization across domains and cameras.
- Hyper-parameter Analysis: As µ increases from 0 to 0.02, rank-1 accuracy rises from 78.6% to 88.8% on Market1501 and from 55.1% to 72.1% on DukeMTMC-ReID.Values of µ greater than 0.01 produce similar performance across datasets.
- Hyper-parameter Analysis: Layer4 AIBN replacement yields larger gains than Layer1 or Layer2, while different α weight-sharing methods have limited impact on performance.Replacing BNs in Layers3 and 4 is selected for the reported experiments.
- Comparison with State-of-the-art Methods: On MSMT17, the method achieves 56.4% rank-1 accuracy, about 11% higher than NRMT despite NRMT using extra DukeMTMC-ReID training data.The method also outperforms existing methods under unsupervised and unsupervised-transfer settings.
- Discussion: A practical limitation is that both stages use pre-defined clustering numbers, making cluster-count selection critical for pseudo-label generation.The paper suggests similarity thresholds as an alternative and leaves generalizable cluster-number strategies for future work.
5. Conclusion
The paper proposes intra-inter camera similarity for unsupervised person ReID, using intra- and inter-camera pseudo-labels to iteratively optimize the model.
- The method trains a multi-branch CNN with intra-camera pseudo-labels, then computes robust inter-camera similarity from classifiers’ scores for cross-camera clustering.