Source-linked AI summary
Range Loss for Deep Face Recognition with Long-tail
Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, Yu Qiao
TL;DR
Long-tailed training data can impair feature learning for rare identities, while cutting data to make class frequencies uniform can omit identity information. The paper proposes range loss to retain and use the full distribution by compacting intra-class features and separating classes, and reports effectiveness and generalizability on LFW and YTF.
Problem
Long-tailed data gives frequent classes greater influence while rare identities have insufficient samples, and uniformizing data can discard information from removed identities.
Method
Range loss jointly supervises CNN learning with softmax loss to reduce intra-class variation and enlarge inter-class distance under imbalanced data.
Results
The method demonstrates effectiveness in overcoming the long-tail effect and generalizability on the LFW and YTF face-recognition benchmarks.
Takeaways & Limitations
Properly tailored long-tailed data can contribute to model training when range loss is combined with softmax loss.
Takeaways & Limitations
Contrastive and triplet losses require sufficient same-person pairs, which are difficult to obtain for rare classes in long-tailed data.
Abstract
from arXiv · showhide
Convolutional neural networks have achieved great improvement on face recognition in recent years because of its extraordinary ability in learning discriminative features of people with different identities. To train such a well-designed deep network, tremendous amounts of data is indispensable. Long tail distribution specifically refers to the fact that a small number of generic entities appear frequently while other objects far less existing. Considering the existence of long tail distribution of the real world data, large but uniform distributed data are usually hard to retrieve. Empirical experiences and analysis show that classes with more samples will pose greater impact on the feature learning process and inversely cripple the whole models feature extracting ability on tail part data. Contrary to most of the existing works that alleviate this problem by simply cutting the tailed data for uniform distributions across the classes, this paper proposes a new loss function called range loss to effectively utilize the whole long tailed data in training process. More specifically, range loss is designed to reduce overall intra-personal variations while enlarging inter-personal differences within one mini-batch simultaneously when facing even extremely unbalanced data. The optimization objective of range loss is the $k$ greatest range's harmonic mean values in one class and the shortest inter-class distance within one batch. Extensive experiments on two famous and challenging face recognition benchmarks (Labeled Faces in the Wild (LFW) and YouTube Faces (YTF) not only demonstrate the effectiveness of the proposed approach in overcoming the long tail effect but also show the good generalization ability of the proposed approach.
1. Introduction
Long-tailed face data gives frequent identities disproportionate influence while leaving rare identities poorly represented. The paper proposes range loss to use the full distribution by reducing intra-class variation and enlarging inter-class differences, with reported gains on LFW and YTF.
- Motivation: Long-tailed data contains few frequent classes and many rare classes, making robust feature learning difficult for poorly sampled identities.Insufficient samples leave intra-class features dispersed while compacting inter-class dispersion.
- Motivation: Uniformizing data by cutting samples can improve performance but may discard information contained in removed identities.The paper contrasts this disposal strategy with retaining the full long-tailed dataset.
- Proposed approach: Range loss minimizes each class’s maximum within-class Euclidean range during training and recomputes the range for the current batch.It is designed to enhance learning for tailed data and identities.
- Contributions: The paper presents range loss as a new loss for addressing long-tail effects in deep face recognition and identifies this problem as previously unaddressed in the literature.This claim is stated as being to the authors’ best knowledge.
- Contributions: Experiments report effectiveness against the long-tail effect and generalizability on the LFW and YTF face-recognition benchmarks.The contribution statement describes both outcomes without specifying additional metrics.
2. Related Work
Related work situates range loss among deep vision models, center loss, and prior studies of long-tailed data. Earlier approaches studied or altered imbalanced data, while center loss provides a related feature-compaction objective.
- Deep vision models: Deep learning methods have achieved strong results in object detection, face recognition, and other vision tasks.The related-work discussion cites VGG, residual networks, and multiple face-recognition systems.
- Dataset construction: The paper constructs long-tailed subsets by dividing object classes according to specified proportions.The division proportions are represented by cutting lines in the constructed-data figure.
- Face-recognition losses: Center loss minimizes distances between deep features and their corresponding arithmetic-mean class centers for face recognition.The paper describes this as similar to range loss’s main practice.
- Long-tailed learning: Long-tailed distributions have been studied in scene parsing and zero-shot learning, with prior work expanding poor classes or analyzing factors affecting detection under imbalance.These studies motivate treating long-tail structure as a broader machine-learning issue.
3. The Proposed Approach
The paper examines how long-tailed training data affects face-recognition features and develops range loss to use such data without discarding samples. Range loss jointly reduces intra-personal variation and increases inter-personal separation using class-level distances within a minibatch.
- Problem formulation: Long-tailed data leave poor identities with few images, producing dispersed intra-class features and compact inter-class features.
- Study of contrastive and triplet losses: Contrastive and triplet losses require sufficient same-person pairs, which are unavailable when tail identities have very few images.
- The Range Loss: Range loss uses all samples in a minibatch jointly, replacing individual pair-based optimization with class-level intra- and inter-class distance statistics.
- The Range Loss: The loss minimizes the harmonic mean of each class’s k greatest Euclidean ranges while maximizing separation through the shortest inter-class center distance.
4. Experiments
Experiments evaluate range loss on LFW and YTF using VGG and residual CNN models trained with long-tailed data. Range loss combined with softmax improves verification accuracy over softmax-only and contrastive-loss baselines.
- Experimental setup: The experiments evaluate range loss on LFW and YTF face-verification benchmarks using VGG and residual CNN architectures.The study trains models with 50% and 100% long-tailed data and compares multiple loss configurations.
- Experimental setup: The training data comprise 5M images from 100k identities after filtering and cleaning MS-Celeb-1M and CASIA-WebFace.Images are detected, aligned, and cropped to 224×224.
- Results: Model C, jointly supervised by range loss and softmax, improves LFW verification accuracy from 97.87% to 98.53% over softmax-only Model A.Table 5 compares softmax-only, contrastive-plus-softmax, and range-loss-plus-softmax models.
- Results: On complete long-tailed data, Model C exceeds the 50% long-tail group’s result by 0.43%.The authors report that range loss enables the latter 50% tailed data to contribute to learning.
- Results: The residual-CNN experiments compare softmax-only Model D with range-loss Model E on LFW and YTF.The comparison is defined in Table 6, while the supplied passages do not include its numerical accuracies.
5. Conclusions
The paper concludes that range loss addresses long-tailed training in face recognition by jointly reducing intra-class variation and enlarging inter-class distance. Experiments on large-scale face benchmarks support its effectiveness.
- 5. Conclusions: Range loss combined with softmax reduces intra-class variations and enlarges inter-class distances under imbalanced long-tailed data.The conclusion emphasizes joint supervision rather than discarding tailed samples.
- 5. Conclusions: The paper reports that appropriately handled long-tailed data can contribute to model training rather than merely harming performance.This conclusion motivates focusing optimization on difficult samples within poor classes.
- 5. Conclusions: Performance on several large-scale face benchmarks is reported as evidence for the effectiveness of the proposed approach.