PaperlayerView exact revision on arXiv

Source-linked AI summary

SphereReID: Deep Hypersphere Manifold Embedding for Person Re-Identification

Xing Fan, Wei Jiang, Hao Luo, Mengjuan Fei

arXiv:1807.00537v1cs.CV

TL;DR

Person ReID commonly uses softmax classification while leaving the embedding-space distribution insufficiently constrained. The paper introduces Sphere Loss and SphereReID, combining hyperspherical embedding with balanced sampling and warm-up training; the resulting single model outperforms state-of-the-art methods on four datasets without fine-tuning or re-ranking.

  • Problem

    Current softmax-based ReID methods classify person images while leaving the underlying feature embedding space insufficiently constrained.

  • Method

    SphereReID uses a modified softmax loss with feature and weight normalization to classify persons while learning embeddings on a hypersphere, with balanced sampling and warm-up training.

  • Results

    94.4% rank-1 accuracy is reported on Market-1501, and the single model outperforms state-of-the-art methods on all four datasets without fine-tuning or re-ranking.

  • Takeaways & Limitations

    The paper presents deep hypersphere manifold embedding as an effective approach for person re-identification.

Abstract

from arXiv · show

Many current successful Person Re-Identification(ReID) methods train a model with the softmax loss function to classify images of different persons and obtain the feature vectors at the same time. However, the underlying feature embedding space is ignored. In this paper, we use a modified softmax function, termed Sphere Softmax, to solve the classification problem and learn a hypersphere manifold embedding simultaneously. A balanced sampling strategy is also introduced. Finally, we propose a convolutional neural network called SphereReID adopting Sphere Softmax and training a single model end-to-end with a new warming-up learning rate schedule on four challenging datasets including Market-1501, DukeMTMC-reID, CHHK-03, and CUHK-SYSU. Experimental results demonstrate that this single model outperforms the state-of-the-art methods on all four datasets without fine-tuning or re-ranking. For example, it achieves 94.4% rank-1 accuracy on Market-1501 and 83.9% rank-1 accuracy on DukeMTMC-reID. The code and trained weights of our model will be released.

1 Introduction

Person ReID methods often use classification or metric learning, but conventional embeddings lack explicit distribution constraints and can suffer from sampling bias. SphereReID addresses these issues with Sphere Loss, balanced sampling, and warm-up training, achieving strong results across four datasets.

  • Person re-identification retrieves gallery images of the same person ID as a probe across non-overlapping camera views.
  • CNN-based ReID commonly treats each person ID as a class and extracts features before the final fully connected layer.
  • Conventional softmax embeddings lack explicit feature-space constraints, producing a generally spread distribution.
  • Triplet loss pulls same-person instances together and pushes different-person instances apart, but embeds features in unbounded Euclidean coordinates.
  • 94.4% rank-1 accuracy is achieved on Market-1501, while a single end-to-end model outperforms state-of-the-art methods on all four datasets without fine-tuning.The reported datasets are Market-1501, DukeMTMC-reID, CUHK03, and CUHK-SYSU.
  • Sphere Loss jointly classifies person images and constrains their embeddings to a hypersphere manifold through a modified softmax objective.

2 Related Works

Related work spans handcrafted and CNN-learned features, supervised metric learning, and loss-function research. Recent face-recognition losses motivate angularly discriminative and normalized embedding objectives relevant to Sphere Loss.

  • Handcrafted ReID representations include SIFT, LBP, and LOMO features, while deep learning automatically learns feature representations from annotated data.
  • Traditional metric-learning methods such as KISSME and XQDA learn feature transformation matrices for person re-identification.
  • Other ReID approaches improve learning through domain-guided dropout, pairwise-consistent dropout, body-part alignment, or teacher-model distillation.
  • Face-recognition research introduced large-margin and angular softmax losses to encourage angularly discriminative features through normalized representations.

3 Our Approach

SphereReID combines Sphere Loss, balanced sampling, and a ResNet-50-based network to learn person embeddings on a hypersphere while performing classification. Sphere Loss removes norm-based decision effects so classification depends on angular relationships.

  • Softmax Loss: Without normalization, both feature and weight norms influence classification, producing weight and feature bias.
  • Sphere Loss: L2-normalizing weights and features removes norm influence, yielding a clear decision boundary determined only by angle.
  • Sphere Loss: The loss adds a scale factor to normalized softmax; the paper uses s = 14 in all experiments.
  • Sphere Loss: Sphere Loss jointly classifies person images and learns an embedding on a hypersphere manifold.
  • Balanced Sampling: Balanced sampling selects P persons and K images per person, forming PK-image mini-batches and equalizing person-instance representation.
  • SphereReID Network: SphereReID uses ResNet-50 followed by global average pooling, batch normalization, dropout, fully connected layers, and L2 normalization.

4 Experiments

The experiments evaluate SphereReID on four widely used person re-identification datasets with varied sizes, identities, camera views, and bounding-box sources.

  • Datasets: Market-1501 contains 32,668 bounding-box images of 1,501 persons, with 751 persons used for training.
  • Datasets: DukeMTMC-reID contains 36,411 images of 1,812 identities across eight cameras, with 702 identities in training and 702 in testing.
  • Datasets: CUHK03 contains 14,096 images of 1,467 identities, and this paper uses its manually cropped bounding-box version.
  • Datasets: CUHK-SYSU contains 99,809 annotated bounding-box images from 18,184 full images and 8,432 identities.

4.2 Implementation Details

SphereReID uses a ResNet-50 backbone with balanced sampling, Adam optimization, and a warming-up learning-rate schedule. Training begins with a gradual learning-rate increase before scheduled decay.

  • The model uses ImageNet-pretrained ResNet-50 after discarding its original fully connected layer.
  • Inputs are resized to 288 × 144 and randomly cropped to 256 × 128, with P = 16 and K = 4 producing mini-batches of 64 images.
  • Adam optimization starts at a learning rate of 10^-3, decays to 10^-4 at epoch 80 and 10^-5 at epoch 100, and runs for 140 epochs.
  • The warming-up schedule linearly increases the learning rate from 5×10^-5 to 10^-3 over 20 epochs.
  • Figure 5 contrasts the warming-up learning-rate schedule with the schedule without warming-up.

4.3 Results of SphereReID

Experiments compare rank-1 accuracy across four ReID datasets and evaluate network structures, balanced sampling, and warming-up. Network-D performs best among the tested structures, while balanced sampling and warming-up improve performance.

  • Experiments compare rank-1 accuracy on Market-1501, DukeMTMC-reID, CUHK03, and CUHK-SYSU.
  • Network structures and loss: Network-D achieves 93.1% rank-1 accuracy on Market-1501 and combines batch normalization with dropout.
  • Network structures and loss: Network-B outperforms network-A, and network-C outperforms network-B, indicating benefits from the additional fully connected layer and batch normalization.
  • Network structures and loss: Sphere Loss performs better than softmax in the reported network-structure experiments.
  • Balanced sampling: Balanced sampling significantly boosts final performance while keeping the network structure unchanged and adding no computational overhead.
  • Warming-up: Warming-up helps the network converge to a better point than applying a large learning rate from the beginning.

Influence of Warming-up.

The experiments compare networks using different dropout ratios, including no dropout and excessive dropout, to assess their influence on performance.

  • A modest dropout ratio of 0.25 outperforms both no dropout and networks using too much dropout.
  • The subsequent experiments fix the dropout ratio at 0.25.
  • Table 3 reports the influence of dropout ratio on network performance.

Ratio of Dropout.

The paper selects dropout and bias settings for SphereReID and examines how test image size affects performance across datasets. The chosen model uses a modest dropout ratio and a learned bias term.

  • Ratio of dropout: A dropout ratio of 0.25 outperforms both no dropout and excessive dropout, so it is fixed for later experiments.
  • Bias term: A network with an automatically learned bias term performs slightly better than one without that bias term.
  • Bias term: The best setting with a learned bias term is subsequently called the SphereReID network.
  • Test image size: Larger test images improve performance on Market-1501, CUHK03, and CUHK-SYSU but worsen it on DukeMTMC-reID.
  • Test image size: The paper attributes DukeMTMC-reID's different trend to its larger border-background area and uses 256 × 128 testing there.

Test Image Size.

The experiment examines how test image size affects performance, comparing resized images with and without center cropping.

  • Images are resized to 288 × 144 with and without center crop of 256 × 128.

4.4 Comparison with the State of the Art

SphereReID consistently outperforms state-of-the-art methods in accuracy and mAP across the evaluated datasets, using a single model without extra attributes, skeleton datasets, or models. It also achieves competitive performance with end-to-end training, no fine-tuning, and a smaller feature size than PCB+RPP.

  • Cross-dataset comparison: 94.4% rank-1 accuracy on Market-1501 is achieved by SphereReID, which consistently outperforms the state of the art in accuracy and mAP.The comparison covers Market-1501, CUHK-SYSU, DukeMTMC-reID, and CUHK03.
  • Model setting: SphereReID uses no extra attributes, skeleton datasets, or models in its network.
  • Training and comparison: SphereReID is trained end-to-end without fine-tuning, unlike PCB+RPP’s three-stage process with fine-tuning.The passage reports that SphereReID is clearly better than PCB+RPP on Market-1501.
  • Feature representation: 1,024-dimensional SphereReID features are compared with PCB+RPP’s 12,288-dimensional features.The authors state that SphereReID features mapped onto a hypersphere manifold are more discriminative.

5 Conclusions

The paper introduces Sphere Loss and SphereReID to learn person-ReID embeddings on a hypersphere manifold. End-to-end training with balanced sampling and warming-up achieves state-of-the-art performance, while the approach offers a new direction for further ReID improvements and broader task exploration.

  • Conclusions: Sphere Loss combines weight normalization and feature normalization in a modified softmax loss for hypersphere manifold embedding.
  • Conclusions: SphereReID trains end-to-end with balanced sampling and warming-up, outperforming the state of the art on all four datasets without re-ranking or fine-tuning.
  • Conclusions: The authors identify deep hypersphere manifold embedding as a new idea for person re-identification and suggest adding a margin term as a future improvement.The proposed margin is intended to increase inter-class variation and reduce intra-class variation.
  • Conclusions: The warming-up strategy can boost deep neural network performance without extra computing overhead and may extend beyond person re-identification.The authors describe it as simple to implement and applicable to other tasks as a direction for future exploration.
Loading 1807.00537v1…