Source-linked AI summary
Deep Representation Learning on Long-tailed Data: A Learnable Embedding Augmentation Perspective
Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, Wenhui Li
TL;DR
Long-tailed training yields narrow tail-class feature distributions that distort the feature space and weaken discriminability. The paper transfers head-class intra-class angular variation to tail classes by constructing feature clouds, with consistent improvements reported on person re-identification and face recognition. The approach also learns distributions adaptively without explicitly defining head and tail classes.
Problem
Long-tailed data give tail classes limited intra-class diversity, producing narrow feature distributions that distort the feature space and compromise discriminative feature learning.
Method
The method transfers head-class intra-class angular distributions to tail classes by constructing feature clouds around tail instances and learning distributions from dataset statistics.
Results
Experiments on person re-identification and face recognition consistently show improvement with the proposed method.
Takeaways & Limitations
Adaptive transfer of head-class intra-class diversity provides a flexible approach to compensating for tail-class diversity during long-tailed feature learning.
Takeaways & Limitations
The framework assumes angular distributions follow Gaussian forms and treats the relevant distribution parameters as independent.
Abstract
from arXiv · showhide
This paper considers learning deep features from long-tailed data. We observe that in the deep feature space, the head classes and the tail classes present different distribution patterns. The head classes have a relatively large spatial span, while the tail classes have significantly small spatial span, due to the lack of intra-class diversity. This uneven distribution between head and tail classes distorts the overall feature space, which compromises the discriminative ability of the learned features. Intuitively, we seek to expand the distribution of the tail classes by transferring from the head classes, so as to alleviate the distortion of the feature space. To this end, we propose to construct each feature into a "feature cloud". If a sample belongs to a tail class, the corresponding feature cloud will have relatively large distribution range, in compensation to its lack of diversity. It allows each tail sample to push the samples from other classes far away, recovering the intra-class diversity of tail classes. Extensive experimental evaluations on person re-identification and face recognition tasks confirm the effectiveness of our method.
1. Introduction
Long-tailed datasets give head classes many samples but tail classes few, producing narrower tail feature distributions that distort the feature space. LEAP transfers head-class intra-class variation to tail classes through learnable feature clouds, and experiments report effectiveness on two recognition tasks.
- Long-tailed datasets contain identities with sufficient samples alongside identities with very few samples, and deep networks perform poorly on tail classes.
- Tail samples occupy a narrow feature-space region because limited data reduce intra-class diversity, compromising their separation from other classes.
- LEAP transfers head-class intra-class angular distributions to tail classes to compensate for their limited diversity during training.
- The method constructs a feature cloud around each tail instance, expanding its distribution range and separating tail classes from others.
- LEAP abandons explicit head-tail definitions and instead learns intra-class distributions adaptively from the dataset distribution.
- Ablation experiments on person re-identification and face recognition demonstrate the effectiveness of the proposed method.
2. Related Work
Prior approaches to imbalanced feature learning use resampling, reweighting, or data augmentation, while loss-function research targets feature discrimination through class-center constraints and angular or cosine margins.
- Imbalanced-data feature learning methods are mainly categorized as resampling, reweighting, and data augmentation.
- Oversampling repeats tail data but may overfit, motivating SMOTE to synthesize samples between tail points and their nearest neighbors.
- Softmax loss mainly checks classification correctness and lacks explicit constraints on inter-class and intra-class distances.
- Angular-margin losses improve feature discrimination by normalizing representations and inserting cosine or angular margins between classes.
3. The Proposed Approach
The method models intra-class feature distributions through angles to class centers, then constructs feature clouds around tail instances by transferring angular variance from better-sampled classes. It combines these reconstructed tail features with original head features for loss computation, with a full version that derives cloud variance from dataset statistics.
- Framework: The framework extracts features, computes class centers, models feature-to-center angles, and uses those distributions to construct tail feature clouds.An angle memory stores class-specific angles, whose means and variances define Gaussian angular distributions.
- Training objective: The method uses CosFace or ArcFace losses to impose angular margins and improve intra-class compactness and inter-class separation during training.The final loss uses original head features together with reconstructed tail features.
- Feature-cloud construction: A feature cloud generates virtual vectors around each tail feature using a distribution learned from head classes, while preserving the tail instance’s identity.The transferred angular variance increases the represented intra-class angular diversity of tail data.
- Vanilla version: In the vanilla version, classes are divided by threshold T: classes with more than T samples are head classes, and the rest are tail classes.This explicit division is described as less flexible than the full version.
- Full version: The full version discards explicit head-tail labels and weights class angular variances by class sample counts to determine the overall variance.Classes with fewer samples contribute less to the final variance, so sufficiently sampled classes dominate the estimate.
4. Experiments
Experiments evaluate LEAP on long-tailed person re-identification and face recognition, using multiple datasets, class-imbalance settings, baselines, and ablations. Results show consistent gains over baselines, flexibility of the full version, and robustness across feature-cloud timings.
- Experimental settings: Experiments cover person re-identification on Market-1501, DukeMTMC-reID, and MSMT17, plus face recognition on MS-Celeb-1M.Long-tailed training sets vary head-class counts and tail samples per class.
- State-of-the-art comparison: The baseline achieves competitive performance, while the full LEAP version further improves results over baseline on Market-1501 and DukeMTMC-reID.The comparison includes state-of-the-art approaches and reports LEPA-CF rank-1 results of 94.4% on Market-1501 and 87.8% on DukeMTMC-reID.
- Vanilla-version evaluation: +6.7% Rank-1 accuracy is achieved by LEAP-CV on DukeMTMC-reID under ⟨H20, S5⟩, increasing performance from 66.0% to 72.7%.LEAP combined with CosFace and ArcFace consistently outperforms the corresponding baselines across different long-tailed settings.
- Vanilla-versus-full comparison: The full version performs close to, and sometimes better than, the vanilla version across long-tailed settings.This supports using the full version without explicitly labeling head and tail classes.
- Impact of tail data: +11.7% Rank-1 accuracy is achieved by LEAP-CF on Market-1501 under ⟨H20, S3⟩, increasing performance from 55.5% to 67.2%.The improvement remains large when each tail class has only three samples.
5. Conclusions
The method transfers intra-class diversity from head classes to tail classes through feature clouds and adaptively learns the intra-class distribution. Experiments on person re-identification and face recognition consistently show improvement.
- The approach transfers intra-class diversity learned from head classes to tail classes during feature learning.
- It constructs a feature cloud for each tail instance during training.
- The method adaptively learns intra-class distributions without explicitly distinguishing head and tail classes.
- Experiments on person re-identification and face recognition consistently show improvement.