PaperlayerView exact revision on arXiv

Source-linked AI summary

Unsupervised Tracklet Person Re-Identification

Minxian Li, Xiatian Zhu, Shaogang Gong

arXiv:1903.00535v1cs.CV

TL;DR

Person re-identification commonly depends on exhaustive per-camera-pair identity labels, limiting scalability and weakening unsupervised accuracy. UTAL learns jointly from automatically extracted, unlabelled tracklets through within-camera discrimination and cross-camera association, outperforming competing methods while requiring sufficient tracklet data.

  • Problem

    Supervised re-identification requires extensive manually labelled cross-view pairs, while unsupervised methods avoid these labels but generally learn weaker discriminative information across cameras.

  • Method

    UTAL jointly learns per-camera tracklet discrimination and cross-camera tracklet association from automatically generated, unlabelled tracklets using shared multi-task features and soft tracklet labels.

  • Results

    UTAL outperformed state-of-the-art competitors across eight benchmark datasets, including MARS gains of 3.1% Rank-1 and 13.8% mAP over DAL.

  • Takeaways & Limitations

    End-to-end learning from unlabelled tracklets makes UTAL more scalable for real-world deployments than approaches requiring exhaustive camera-pair or source-domain labels.

  • Takeaways & Limitations

    UTAL requires sufficient tracklet data: it is inferior to top existing models on the small iLIDS-VID and PRID2011 benchmarks.

Abstract

from arXiv · show

Most existing person re-identification (re-id) methods rely on supervised model learning on per-camera-pair manually labelled pairwise training data. This leads to poor scalability in a practical re-id deployment, due to the lack of exhaustive identity labelling of positive and negative image pairs for every camera-pair. In this work, we present an unsupervised re-id deep learning approach. It is capable of incrementally discovering and exploiting the underlying re-id discriminative information from automatically generated person tracklet data end-to-end. We formulate an Unsupervised Tracklet Association Learning (UTAL) framework. This is by jointly learning within-camera tracklet discrimination and cross-camera tracklet association in order to maximise the discovery of tracklet identity matching both within and across camera views. Extensive experiments demonstrate the superiority of the proposed model over the state-of-the-art unsupervised learning and domain adaptation person re-id methods on eight benchmarking datasets.

1 INTRODUCTION

The paper addresses the scalability limits of supervised person re-identification by learning from automatically generated tracklets without identity labels. UTAL jointly learns within-camera discrimination and cross-camera association, and experiments compare it favorably with unsupervised and domain-adaptation methods.

  • Motivation: Supervised re-identification requires extensive manually labeled cross-view pairs for each camera pair, limiting scalability across camera networks.The required labels support camera-pair-specific feature or distance learning.
  • Motivation: Unsupervised re-identification avoids per-camera-pair labels but has lower accuracy because missing cross-view supervision weakens discriminative learning.This is especially problematic under substantial appearance changes across cameras.
  • Motivation: Domain-adaptation approaches use unlabeled target data with labeled source data, assuming shared cross-view characteristics that may not hold.The paper instead considers a pure unsupervised setting with no labeled source, cross-view, or within-view identity data.
  • Approach: UTAL jointly performs per-camera tracklet discrimination and cross-camera tracklet association using automatically generated tracklets without camera-pair identity labels.Its two learning components target local within-camera label discrimination and global cross-view identity matching.
  • Evaluation: Experiments span eight datasets and report performance advantages over state-of-the-art unsupervised and domain-adaptation re-identification models.The evaluation includes seven existing benchmarks and the newly introduced DukeTracklet dataset.
  • Advances: Compared with an earlier version, the study learns directly from large-scale raw tracklets, adds soft tracklet labeling, and improves end-to-end handling of trajectory fragmentation.The revised method avoids domain-specific pre-sampling and retains potentially informative tracklets.

2 RELATED WORK

Related work spans supervised, semi-supervised, domain-adaptation, and unsupervised re-identification, but these approaches retain labeling, domain, or clustering constraints. UTAL instead learns from automatically generated tracklets across unconstrained camera domains using task-level relationships without sample-wise multi-task labels.

  • Re-identification learning: Supervised re-identification depends on per-camera-pair identity labels, while semi-supervised methods still require substantial cross-view labeled data.Unsupervised learning improves scalability by avoiding exhaustive domain-specific labeling, but earlier hand-crafted approaches have inferior matching performance.
  • Domain adaptation: Domain adaptation transfers knowledge from labeled source domains to unlabeled targets, often relying on related-domain assumptions or explicit label-knowledge transfer.Such assumptions and transfer procedures constrain applicability when source and target scenes differ.
  • UTAL: UTAL learns re-identification features from automatically generated unlabeled tracklets end-to-end without source labels, camera-domain similarity assumptions, or cross-view identity overlap.The method is designed for unconstrained viewing settings and does not require camera space-time topology.
  • Multi-task learning: Unlike conventional multi-task learning, UTAL correlates group-level semantic relationships across camera-specific tasks rather than relying on clean sample-wise labels.Sample-wise multi-task labels are unavailable in the unsupervised tracklet setting.
  • Unsupervised representation learning: UTAL’s soft-label affinity approach relates to clustering while avoiding per-domain hard clustering and the need to tune camera-specific cluster counts.This design supports learning across different camera-domain distributions.

3 METHOD FORMULATION

UTAL learns person re-identification representations directly from automatically generated, unlabeled tracklets by jointly optimizing within-camera discrimination and cross-camera association. Its end-to-end design combines multi-task per-camera learning, soft labels for fragmented trajectories, and progressively discovered cross-camera matches.

  • 3.1 Unsupervised Per-Camera Tracklet Formation: UTAL uses automatically generated tracklets without manual identity verification or within-camera or cross-camera ID labels.Tracklets are formed independently for each camera from detection and tracking outputs, with each tracklet assigned an unsupervised class label.
  • 3.2 Unsupervised Tracklet Association: The framework jointly optimizes Per-Camera Tracklet Discrimination and Cross-Camera Tracklet Association in one end-to-end neural architecture.The two components are designed to mutually benefit representation learning during incremental optimization.
  • 3.2.1 Per-Camera Tracklet Discrimination Learning: PCTD treats each camera as a separate classification task while sharing a common feature representation across all camera-specific branches.With T cameras, UTAL forms T tracklet-classification tasks and optimizes their cross-entropy losses in a multi-task network.
  • 3.2.1 Per-Camera Tracklet Discrimination Learning: The shared representation is updated efficiently using tracklet feature vectors maintained across iterations instead of forwarding all tracklet data each time.The external tracklet representations are progressively updated with image features, and the learning rate controls their update speed.
  • 3.2.1 Per-Camera Tracklet Discrimination Learning: Soft tracklet labels use within-camera appearance affinity to improve robustness when trajectory fragmentation assigns distinct labels to the same identity.This replaces potentially misleading hard labels and indirectly refines fragmented tracklets.
  • 3.2.2 Cross-Camera Tracklet Association Learning: PCTD alone lacks explicit fine-grained cross-camera association constraints, motivating CCTA to discover informative cross-view matches.Most cross-camera pairs are unmatched identities, so the method must identify positive associations despite unreliable representations and possible false-positive propagation.
  • 3.2.2 Cross-Camera Tracklet Association Learning: CCTA grounds cross-camera learning on nearest-neighbour tracklets and incrementally adds constraints only for discovered cross-camera matches.Its applicability is model-maturity adaptive because the set of matches can grow as training improves the representation.
  • 3.2.3 Joint Unsupervised Tracklet Association Learning: The final UTAL objective combines the image-level PCTD loss with the tracklet-group-level CCTA loss using a balance weight λ.Both losses are derived concurrently from the same mini-batch, integrating local discrimination with cross-camera association.

4 EXPERIMENTS

Experiments evaluate UTAL on image, video, and tracklet benchmarks, comparing its components, variants, and state-of-the-art unsupervised methods. Results show strong performance when sufficient tracklet data are available, with measurable benefits from PCTD, soft labels, and CCTA.

  • Experimental Setting: UTAL is evaluated on seven established image/video datasets and the newly introduced DukeTracklet benchmark using standard test protocols.The evaluation includes CUHK03, Market-1501, DukeMTMC-ReID, MSMT17, iLIDS-VID, PRID2011, MARS, and DukeTracklet.
  • Comparisons to the State-Of-The-Art Methods: UTAL outperforms existing video re-id models on MARS by 3.1% Rank-1 and 13.8% mAP over DAL.The reported margins are 49.9−46.8 Rank-1 and 35.2−21.4 mAP.
  • Comparisons to the State-Of-The-Art Methods: UTAL requires sufficient tracklet data: it outperforms competitors on MARS but is inferior to top models on the smaller iLIDS-VID and PRID2011 benchmarks.MARS has 8,298 training tracklets, compared with 300 for iLIDS-VID and 178 for PRID2011.
  • Component Analysis and Discussion: PCTD improves Rank-1 by 27.9% on MARS and 27.8% on DukeTracklet over the corresponding comparison, supporting its role in exploiting unsupervised tracklet labels.The reported gains are 43.8−15.9 on MARS and 31.7−3.9 on DukeTracklet.
  • Component Analysis and Discussion: Removing CCTA reduces Rank-1 by 6.1% on MARS and 12.1% on DukeTracklet, demonstrating the contribution of cross-camera association learning.The comparison reports drops from 49.9 to 43.8 on MARS and from 43.8 to 31.7 on DukeTracklet.
  • Component Analysis and Discussion: The CCTA matching process improves during training as both the number and precision of self-discovered cross-camera pairs increase.This trend is reported as evidence of progressively stronger cross-camera tracklet matching.
  • Component Analysis and Discussion: UTAL substantially improves over TAUDL and benefits from stronger backbones, while using all tracklets outperforms the SSTT strategy.Rank-1 improves over TAUDL by 6.1% on MARS and 17.7% on DukeTracklet; SSTT causes drops of 4.8% and 12.1%, respectively.

5 CONCLUSIONS

UTAL enables unsupervised person re-identification from totally unlabelled tracklet data by jointly learning per-camera discrimination and cross-camera association. Its unified architecture removes exhaustive pairwise identity labelling requirements and supports scalable real-world deployment.

  • UTAL learns an unsupervised person re-identification model end-to-end from automatically extracted, totally unlabelled tracklet data.This eliminates expensive and exhaustive manual identity labelling.
  • The unified architecture jointly optimises Per-Camera Tracklet Discrimination and Cross-Camera Tracklet Association losses.
  • UTAL is more scalable than methods requiring exhaustive pairwise labels for every camera pair or labelled source-domain training data.
  • Evaluations on eight image and video person re-identification benchmarks validate UTAL against state-of-the-art unsupervised and domain adaptation methods.
Loading 1903.00535v1…