PaperlayerView exact revision on arXiv

Source-linked AI summary

ICE: Inter-instance Contrastive Encoding for Unsupervised Person Re-identification

Hao Chen, Benoit Lagadec, Francois Bremond

arXiv:2103.16364v2cs.CV

TL;DR

Unsupervised ReID needs discriminative identity features without costly annotations, while prior contrastive methods underexplore relationships between different instances and can suffer from pseudo-label noise. ICE uses pairwise similarity ranking for hard instance contrast and similarity scores for augmentation consistency. It significantly outperforms previous unsupervised ReID methods on Market1501, DukeMTMC-reID, and MSMT17.

  • Problem

    Prior contrastive ReID methods do not fully explore inter-instance relationships, while hardest-positive mining and one-hot labels can introduce false positives or overlook complex pseudo-instance relationships.

  • Method

    ICE combines centroid contrast with pairwise hard labels for hardest-positive contrast and pairwise soft labels for consistency between augmented and original instances.

  • Results

    ICE significantly outperforms previous unsupervised ReID methods on Market1501, DukeMTMC-reID, and MSMT17.

  • Takeaways & Limitations

    ICE demonstrates the effectiveness of inter-instance pairwise similarity for unsupervised person ReID across three mainstream datasets.

  • Takeaways & Limitations

    Without ground truth, mining hardest positives within clusters can introduce false positives during training.

Abstract

from arXiv · show

Unsupervised person re-identification (ReID) aims at learning discriminative identity features without annotations. Recently, self-supervised contrastive learning has gained increasing attention for its effectiveness in unsupervised representation learning. The main idea of instance contrastive learning is to match a same instance in different augmented views. However, the relationship between different instances has not been fully explored in previous contrastive methods, especially for instance-level contrastive loss. To address this issue, we propose Inter-instance Contrastive Encoding (ICE) that leverages inter-instance pairwise similarity scores to boost previous class-level contrastive ReID methods. We first use pairwise similarity ranking as one-hot hard pseudo labels for hard instance contrast, which aims at reducing intra-class variance. Then, we use similarity scores as soft pseudo labels to enhance the consistency between augmented and original views, which makes our model more robust to augmentation perturbations. Experiments on several large-scale person ReID datasets validate the effectiveness of our proposed unsupervised method ICE, which is competitive with even supervised methods. Code is made available at https://github.com/chenhao2345/ICE.

1. Introduction

Unsupervised ReID avoids costly cross-camera identity annotation, but prior contrastive methods underuse inter-instance relationships and can mishandle pseudo-label uncertainty. ICE addresses these issues with hard and soft pairwise similarity supervision within a fully unsupervised framework.

  • Motivation: Unsupervised ReID trains on unlabeled target data, improving scalability when cross-camera identity annotation or large labeled source datasets are unavailable.UDA performance depends strongly on source-dataset scale and quality, whereas fully unsupervised methods require no identity annotation.
  • Motivation: Prior instance contrastive methods treat each image as a separate class, while SpCL uses cluster centroids but neglects inter-instance affinities.This is problematic for fine-grained ReID because multiple images of one identity may be treated as different classes.
  • Motivation: Hardest-positive mining can introduce false positives without ground truth, and one-hot pseudo-labels overlook complex relationships among multiple pseudo positives and negatives.These limitations motivate combining hard pairwise labels with similarity-based soft labels.
  • ICE: ICE uses pairwise similarity scores as soft pseudo labels to enforce consistency between augmented and original instances, improving robustness to augmentation perturbations.The soft labels account for similarity changes caused by occlusion, viewpoint, and resolution distortions.
  • ICE: Without identity annotation, ICE combines centroid contrast, hardest-positive contrast, and augmentation consistency and significantly outperforms state-of-the-art UDA and fully unsupervised methods.The comparison is reported on mainstream person ReID datasets.
  • ICE: ICE ranks pairwise similarities to form one-hot hard pseudo labels for hardest-positive contrast, reducing intra-class variance.The hard instance objective emphasizes difficult samples within clusters rather than relying only on class-level centroids.

2. Related Work

Prior unsupervised ReID work includes UDA, pseudo-labeling, and centroid-based contrastive approaches, while contrastive learning and consistency regularization provide complementary foundations. ICE is positioned as combining intra-class variance reduction with consistency between augmented and original views.

  • Unsupervised person ReID: Unsupervised ReID methods include UDA approaches using attributes, image generation, clustering, reference data, or temporal information to adapt or label unlabeled images.These approaches are part of a broader taxonomy that also includes fully unsupervised methods.
  • Unsupervised person ReID: SpCL and CAP perform instance-to-centroid contrast but neglect inter-instance affinities.SpCL treats clusters and outliers as classes, whereas CAP uses camera-specific identity centroids.
  • Contrastive Learning: Memory banks, momentum encoders, and large batches support contrastive representation learning, but standard methods treat different instances of one class as different classes.That assumption is unsuitable for fine-grained ReID, where multiple images can share an identity.
  • Consistency regularization: Consistency regularization assumes predictions should remain consistent under perturbed versions of the same image.Perturbations can arise from augmentation, temporal ensembling, or shallow-deep features.
  • Consistency regularization: ICE simultaneously reduces intra-class variance and regularizes consistency between augmented and original views for fine-grained ReID.This combines a concern emphasized by contrastive ReID work with a consistency-based response to data variance.

3. Proposed Method

ICE combines class-level proxy contrast with hard and soft inter-instance signals in a fully unsupervised ReID framework. It uses momentum representations, clustering, proxy memories, hardest-positive matching, and augmentation-consistency regularization.

  • Overview: ICE uses an online encoder updated by back-propagation and a momentum encoder updated from accumulated online-encoder weights.The momentum encoder produces more stable representations under label noise.
  • Overview: At each epoch, ICE clusters momentum representations with DBSCAN, discards outliers, stores cluster proxies, and samples mini-batches containing multiple pseudo identities.Training combines proxy contrast, hard instance contrast, and soft instance consistency losses.
  • 3.2. Proxy Centroid Contrastive Baseline: Proxy contrast pulls instances toward cluster centroids and, when camera information is available, toward cross-camera centroids to reduce camera-style variance.The camera-agnostic version uses one positive cluster proxy and negative proxies from memory; the combined proxy loss weights the components using coefficient 0.5.
  • 3.3. Hard Instance Contrastive Loss: The hard instance contrastive loss selects each anchor’s lowest-similarity positive and contrasts it against all negatives in the mini-batch.Minimizing distance to the hardest positive and maximizing distance from negatives increases intra-class compactness and inter-class separability; unlike triplet loss, it pushes all negatives through a softmax.
  • 3.4. Soft Instance Consistency Loss: Soft instance consistency converts augmented inter-instance similarities into a prediction distribution and uses unaugmented similarities as soft labels to preserve relationships across perturbations.The design addresses similarity changes among pseudo positives and negatives after augmentation.

4. Experiments

Experiments evaluate ICE on three person ReID datasets, analyze its parameters and loss components, and compare it with unsupervised and supervised methods. The results show improved cluster compactness, augmentation robustness, and performance across evaluation settings.

  • Experimental setup: Experiments use Market-1501, DukeMTMC-reID, and MSMT17 datasets under standardized training and clustering settings.The backbone is ImageNet-pretrained ResNet50; pseudo labels are renewed every 400 iterations for 40 epochs, and DBSCAN uses k-reciprocal distance with threshold 0.55.
  • Parameter analysis: Parameter analysis on Market-1501 studies the weights and temperatures of hard instance contrast and soft instance consistency losses.The selected loss weights are λh ins = 1 and λs ins = 10; the passage attributes excessive hard-loss weight or insufficient hard-loss temperature to increased noise from false positives.
  • Ablation study: Without hard instance contrast, DBSCAN generates more clusters during training, whereas full ICE produces fewer, more compact clusters.The authors connect reduced intra-class variance with cluster compactness and improved unsupervised ReID performance.
  • Ablation study: Removing soft instance consistency makes the model less robust to augmentation perturbations, while combining both proposed losses gives the best camera-aware performance.Hard contrast primarily reduces variance between naturally captured views; soft consistency mainly reduces variance from artificial augmentation.
  • Comparison with prior methods: ICE outperforms state-of-the-art UDA and fully unsupervised methods across the three datasets without identity annotation.With camera-agnostic memory it surpasses SpCL, and with camera-aware memory it outperforms CAP on all three datasets.
  • Comparison with supervised methods: Unsupervised ICE achieves competitive performance with the supervised PCB method, while supervised ICE is competitive with state-of-the-art supervised ReID methods.The supervised variant replaces clustering-generated pseudo labels with ground-truth labels.

5. Conclusion

ICE uses inter-instance affinities to address variance in unsupervised ReID. Its hard and soft objectives reduce intra-class and augmentation variance, and the combined method significantly outperforms previous unsupervised methods on three datasets.

  • Conclusion: ICE mines hardest positives with mini-batch pairwise similarity ranking to reduce intra-class variance and improve cluster compactness.The hard instance contrastive loss forms one-hot hard pseudo labels from pairwise similarity rankings.
  • Conclusion: ICE uses mini-batch pairwise similarity scores as soft labels to improve consistency before and after augmentation.The soft instance consistency loss is intended to improve robustness to artificial augmentation variance.
  • Conclusion: ICE significantly outperforms previous unsupervised ReID methods on Market1501, DukeMTMC-reID, and MSMT17.The conclusion reports this result across all three evaluated datasets.

Appendix A. Algorithm Details

Algorithm 1 describes fully unsupervised ICE training from unlabeled images to a trained momentum encoder. The procedure encodes data, clusters momentum representations, computes proxies, and combines three contrastive objectives.

  • Algorithm inputs and output: ICE takes an unlabeled dataset, pretrained online and momentum encoders, and maximum epoch and iteration settings as inputs.The algorithm outputs the momentum encoder after training.
  • Pseudo-label generation: The algorithm encodes the dataset into momentum representations and generates clustering pseudo labels using reranking and DBSCAN.These operations form the unlabeled training structure used by subsequent proxy calculations.
  • Proxy construction: ICE calculates cluster and camera proxies from the generated pseudo labels.The proxies are defined in the cited equations and support the contrastive training objectives.
  • Loss combination: The training objective combines proxy contrastive loss, hard instance contrastive loss, and soft instance consistency loss.These three losses correspond to the class-level, hard pairwise, and soft pairwise components of ICE.

Appendix B. Backbone Network

Replacing the ResNet50 backbone with IBN-ResNet50 further improves ICE performance. This comparison extends the method's evaluation to a backbone designed for stronger normalization behavior.

  • Backbone comparison: ICE performs better with an IBN-ResNet50 backbone than with a regular ResNet50 backbone.The paper compares both backbones in Table 6 and reports further improvement with IBN-ResNet50.

Appendix C. Threshold in clustering

The clustering distance threshold controls which samples DBSCAN treats as neighbors, creating a trade-off between separating hard positives and merging hard negatives. Threshold sensitivity varies across datasets, with smaller values favored for Market1501 and larger values for MSMT17.

  • Threshold effects: DBSCAN’s distance threshold determines the maximum neighbor distance used for clustering.Lower thresholds can separate more hard positives, whereas higher thresholds can merge more hard negatives.
  • Threshold settings: The main-paper threshold is 0.55 for Market1501, DukeMTMC-reID, and MSMT17.This value is described as a trade-off across the three datasets.
  • Cluster compactness: Figure 10 tracks dynamic cluster numbers during 40 training epochs on DukeMTMC-reID.Lower cluster numbers indicate more compact clusters with less intra-cluster variance.
  • Threshold settings: Threshold sensitivity is evaluated from 0.45 to 0.6 across Market1501, DukeMTMC-reID, and MSMT17.Table 7 reports comparisons across these distance-threshold settings.
  • Dataset sensitivity: A threshold of 0.5 is more appropriate for Market1501, while 0.6 is more appropriate for MSMT17.The passage associates the smaller value with the relatively smaller dataset and the larger value with the relatively larger dataset.

Appendix D. Camera-agnostic scenario

The camera-agnostic analysis examines how hard instance contrast and soft instance consistency jointly affect cluster compactness during training. Combining both losses reduces naturally captured and artificially augmented view variance and yields optimal ReID performance.

  • Camera-agnostic analysis: Without hard instance contrastive loss Lh ins, soft instance consistency loss Ls ins maintains high intra-class variance and produces less compact clusters.The analysis describes this condition using approximately equal pairwise similarity values among the anchor and positives.
  • Camera-agnostic analysis: Combining Lh ins and Ls ins reduces both naturally captured and artificially augmented view variance.The joint design is reported to produce optimal ReID performance.

Appendix E. Future work

ICE is designed for traditional short-term person ReID, where people do not change clothes. Clothing changes in long-term ReID can make its visual-similarity-based pseudo labels less robust, motivating future work on more robust pseudo-label generation.

  • Scope limitation: ICE’s scope is traditional short-term person ReID, where persons do not change their clothes.The method is not presented as robust to long-term clothing changes.
  • Future work: In long-term ReID, taking off or changing clothes can produce less robust pseudo labels because they rely mainly on visual similarity.The proposed future direction is generating robust pseudo labels for clothing-change scenarios.
Loading 2103.16364v2…