Source-linked AI summary
Unsupervised Person Re-identification via Multi-label Classification
Dongkai Wang, Shiliang Zhang
TL;DR
Unsupervised person ReID must learn discriminative features without true labels. The paper iteratively predicts multi-class labels with similarity and cycle consistency, then trains with MMCL; it reports competitive performance without labeled data and supports transfer learning.
Problem
Unsupervised person ReID must learn discriminative features without true labels or, in prior approaches, reliance on labeled data or a good pre-trained model.
Method
The method iteratively predicts multi-class labels with MPLP using visual similarity and cycle consistency, then updates the ReID network with the memory-based multi-label classification loss MMCL.
Results
80.3% rank-1 accuracy is achieved on Market-1501 without other labeled data, outperforming BUC by 14.1% and DBC by 11.1%.
Takeaways & Limitations
The iterative MPLP–MMCL strategy demonstrates effective unsupervised person ReID and remains compatible with transfer learning, reaching 84.4% rank-1 accuracy on Market-1501 with DukeMTMC-reID training data.
Takeaways & Limitations
The multi-label loss must account for the imbalance between positive and negative classes because treating negative classes equally may cause model collapse.
Abstract
from arXiv · showhide
The challenge of unsupervised person re-identification (ReID) lies in learning discriminative features without true labels. This paper formulates unsupervised person ReID as a multi-label classification task to progressively seek true labels. Our method starts by assigning each person image with a single-class label, then evolves to multi-label classification by leveraging the updated ReID model for label prediction. The label prediction comprises similarity computation and cycle consistency to ensure the quality of predicted labels. To boost the ReID model training efficiency in multi-label classification, we further propose the memory-based multi-label classification loss (MMCL). MMCL works with memory-based non-parametric classifier and integrates multi-label classification and single-label classification in a unified framework. Our label prediction and MMCL work iteratively and substantially boost the ReID performance. Experiments on several large-scale person ReID datasets demonstrate the superiority of our method in unsupervised person ReID. Our method also allows to use labeled person images in other domains. Under this transfer learning setting, our method also achieves state-of-the-art performance.
1. Introduction
The paper addresses unsupervised person ReID by iteratively predicting multi-class identity labels and training with a memory-based multi-label loss. Without other labeled data, this approach achieves competitive performance across large-scale datasets and also supports transfer learning.
- Unsupervised person ReID seeks discriminative features from unlabeled images because annotating images across multiple cameras is expensive.
- The method iteratively runs MPLP and MMCL to seek true labels, and the maintained memory bank reinforces label prediction and classification.The paper reports that this strategy achieves promising performance without leveraging labeled data and remains compatible with transfer learning.
- The method treats each unlabeled image as a class, then assigns multiple classes to images believed to share an identity.Multi-label classification is intended to identify images of the same person and distinguish images from different people.
- MPLP predicts labels using visual similarity, cycle consistency, and memory-bank features updated with augmented representations.Two images receive the same label when they have high similarity and similar neighbors.
- MMCL uses memory-bank image features as a classifier to accelerate multi-label loss computation and address vanishing gradients.It is designed for the large number of classes created when each image is initially treated as a class.
- 80.3% rank-1 accuracy is achieved on Market-1501 without other labeled data, outperforming BUC by 14.1% and DBC by 11.1%.With DukeMTMC-reID used for training, rank-1 accuracy reaches 84.4% on Market-1501.
2. Related Work
Prior unsupervised person ReID work uses hand-crafted features, clustering, or transfer learning, while this paper applies multi-label classification to learn identity features without labeled data.
- Hand-crafted feature methods are difficult to design for robust and discriminative person ReID representations.
- Clustering methods estimate pseudo labels for CNN training but require a good pre-trained model.
- Transfer-learning methods use source datasets, attributes, soft labels, or generated images to improve unsupervised person ReID.
- Unsupervised feature learning can initialize networks or acquire representations using signals such as image rotation, motion, and viewpoint.
- Multi-label classification has been used for partial-label learning and attribute features, whereas this paper uses it to learn identity features for unsupervised person ReID.The paper describes this as an early use of multi-label classification for unsupervised person ReID.
3. Methodology
The method iteratively converts single-class pseudo-labels into multi-label identity cues using memory-based label prediction, then trains ReID with MMCL. MPLP combines similarity and cycle consistency, while MMCL addresses inefficient and unstable multi-label classification.
- Formulation: Each image initially forms an individual class with a two-valued pseudo-label, which is later expanded into multiple identity-related classes.The initial label sets only image i to 1 and all other classes to −1; multi-label prediction is needed because one person may have multiple images.
- Memory-based framework: The memory bank stores image features and supports both MPLP for label prediction and MMCL for ReID model training.M[i] stores the feature of image xi, and the memory is updated after training iterations with normalized augmented features.
- Memory-based Positive Label Prediction: MPLP assigns multi-label predictions by ranking images using feature similarity, thresholding candidates, and filtering them with cycle consistency.Cycle consistency retains candidates whose neighbor sets are mutually similar, helping reject hard negatives; the procedure adaptively selects candidate counts.
- Memory-based Positive Label Prediction: MPLP consistently outperforms KNN in precision and recall across training stages and effectively rejects hard negative labels.The figure reports both retrieval-quality measures and the selected positive and negative labels.
- Memory-based Multi-label Classification Loss: MMCL abandons sigmoid activation, regresses bounded classification scores toward 1 and −1, and mines hard negative classes to improve efficiency and address class imbalance.Traditional MCL suffers from vanishing gradients and can collapse when numerous negative classes are treated equally; MMCL addresses these issues and is more effective for ReID optimization.
4. Experiments
Experiments evaluate the method on three person ReID datasets, analyze key hyperparameters and components, and compare performance with existing unsupervised and transfer-learning approaches.
- Datasets: Market-1501, DukeMTMC-reID, and MSMT17 provide evaluation settings, with MSMT17 described as the largest and most challenging dataset.Market-1501 has 32,668 images and 1,501 identities; DukeMTMC-reID has 36,411 images and 1,404 identities; MSMT17 has 126,411 images and 4,101 identities.
- Parameter Analysis: The best MPLP similarity threshold is t = 0.6 on both Market-1501 and DukeMTMC-reID.Lower thresholds introduce many negative labels, whereas overly large thresholds select too few labels.
- Parameter Analysis: MMCL performs poorly with δ = 1, reaching 59.3% and 52.6% rank-1 accuracy on Market-1501 and DukeMTMC-reID, respectively; larger values improve performance until training becomes unstable.The experiments therefore set δ = 5.
- Parameter Analysis: Hard negative mining improves performance over using all negative classes, but selecting too few negatives is also harmful; the experiments use r = 1.The analysis identifies a trade-off between retaining informative negatives and avoiding unhelpful easy negatives.
- Ablation Study: Combining MMCL with MPLP raises Market-1501 rank-1 accuracy from the 7.8% baseline to 80.3%, while single-class pseudo-labeling also improves the baseline.The comparison includes supervised learning, MMCL with single-class labels, MMCL with MPLP, and the ImageNet-pretrained baseline.
- Ablation Study: MPLP outperforms KNN and similarity-score label selection, while MMCL outperforms cross-entropy across single-class, ground-truth, and MPLP-predicted label settings.MPLP achieves this by combining similarity score selection with cycle consistency; MMCL remains competitive with supervised learning despite using non-parametric classifiers.
- Comparison with the State of the Art: On MSMT17, the method achieves 35.4% rank-1 accuracy in the unsupervised setting and 43.6%/40.8% under transfer learning, exceeding SSG by 11.4% in rank-1 accuracy.The paper reports large-margin improvements over existing methods under both settings.
5. Conclusion
The paper addresses unsupervised person ReID without labeled data or a good pretrained model through iterative multi-class label prediction and network updates.
- The method performs unsupervised person ReID without requiring labeled data or a good pretrained model.
- It iteratively predicts multi-class labels and updates the network using a multi-label classification loss.
- MPLP predicts multi-class labels using visual similarity and cycle consistency.