PaperlayerView exact revision on arXiv

Source-linked AI summary

A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification

Hao Luo, Wei Jiang, Youzhi Gu, Fuxu Liu, Xingyu Liao, Shenqi Lai, Jianyang Gu

arXiv:1906.08332v2cs.CV

TL;DR

Person ReID research often uses complex structures and understated training tricks, making strong and fair baselines important. This paper collects and evaluates effective tricks, introduces BNNeck to separate classification and metric losses, and combines them in a global-feature ResNet50 baseline. The baseline achieves 94.5% rank-1 accuracy and 85.9% mAP on Market1501, while extended experiments assess its broader applicability.

  • Problem

    Many ReID methods use complex features, and effective training tricks are often understated, while many published approaches build on weak baselines.

  • Method

    The paper evaluates training tricks, proposes BNNeck to place classification and metric losses in separate feature spaces, and adds center loss for clustering properties.

  • Results

    94.5% rank-1 accuracy and 85.9% mAP were achieved on Market1501 by the ResNet50 global-feature baseline.

  • Takeaways & Limitations

    The baseline can serve as a strong foundation for person ReID, and the paper reports that some state-of-the-art methods can be extended with it.

Abstract

from arXiv · show

This study explores a simple but strong baseline for person re-identification (ReID). Person ReID with deep neural networks has progressed and achieved high performance in recent years. However, many state-of-the-art methods design complex network structures and concatenate multi-branch features. In the literature, some effective training tricks briefly appear in several papers or source codes. The present study collects and evaluates these effective training tricks in person ReID. By combining these tricks, the model achieves 94.5% rank-1 and 85.9% mean average precision on Market1501 with only using the global features of ResNet50. The performance surpasses all existing global- and part-based baselines in person ReID. We propose a novel neck structure named as batch normalization neck (BNNeck). BNNeck adds a batch normalization layer after global pooling layer to separate metric and classification losses into two different feature spaces because we observe they are inconsistent in one embedding space. Extended experiments show that BNNeck can boost the baseline, and our baseline can improve the performance of existing state-of-the-art methods. Our codes and models are available at: https://github.com/michuanhaohao/reid-strong-baseline.

I. INTRODUCTION

The paper argues that person ReID needs stronger, fairly evaluated baselines because many methods rely on understated training tricks, complex features, or weak baselines. It collects and evaluates effective tricks, introduces BNNeck and center loss, and builds a strong global-feature baseline.

  • Motivation: Only two of 23 surveyed baselines exceeded 90% rank-1 accuracy on Market1501, while four remained below 80%.The authors use these results to argue that strong baselines are important for demonstrating method effectiveness.
  • Motivation: Understated training tricks can make methods appear more effective and complicate fair comparisons with other state-of-the-art approaches.The paper recommends that reviewers consider such tricks when evaluating academic papers.
  • Motivation: Complex local-feature, pose, and segmentation methods add computation or reduce retrieval speed, motivating a simple model using only global features.The industry motivation is to improve models without extra consumption or concatenated local features.
  • Contributions: The study collects effective person ReID training tricks and evaluates each trick on two widely used datasets.It selects six tricks after numerous experiments.
  • Contributions: BNNeck separates classification and metric losses into different feature spaces because the paper observes inconsistency between ID and triplet losses.The paper also proposes center loss to address neglected intra-class compactness and inter-class separability.
  • Results: 94.5% rank-1 accuracy and 85.9% mAP were achieved on Market1501 by the ResNet50 global-feature baseline.The authors report this as the best performance acquired by global features in person ReID to their knowledge.

II. RELATED WORKS

Related work frames deep person ReID around backbone networks and classification or metric losses, alongside baseline designs that use global or part-level features. It also identifies batch normalization after global pooling as an established baseline practice.

  • Backbones and losses: Deep person ReID commonly uses CNN backbones such as GoogleNet, ResNet, and DenseNet for end-to-end person-image representation.The literature is organized partly by the backbone networks used for feature extraction.
  • Backbones and losses: ReID baselines are broadly categorized by classification loss or metric loss.The related work introduces ID-discriminative embedding for classification-based training and triplet loss for metric learning.
  • Metric learning: Triplet loss treats ReID as a clustering or ranking problem using anchor, positive, and negative images.It reduces anchor-positive distance while increasing anchor-negative distance.
  • Baseline designs: A prior effective CNN baseline adds a batch normalization layer after global pooling, while PCB represents images with several part-level features.Both global- and part-based baselines are described as achieving good person ReID performance.

B. Some Existing Approaches for Deep person ReID

Existing person ReID approaches add local, pose, mask, attention, synthetic, or re-ranking components to improve representations or retrieval. The paper positions its standard baseline and training tricks as a simpler foundation for these methods.

  • Stripe-based methods: Stripe-based methods extract local features from image stripes but can suffer from pose misalignment.Examples include AlignedReID++, MGN, and SCPNet.
  • Pose-guided methods: Pose-guided methods use an extra pose or skeleton model to align corresponding body parts, trading performance against additional computation.The paper identifies model speed as an important consideration.
  • Mask-guided methods: Mask-guided methods use segmentation or semantic parsing to suppress background clutter and exploit local visual cues, but rely heavily on accurate pedestrian masks.Examples include mask-guided contrastive attention and SPReID.
  • Attention-based methods: Attention-based methods learn additional discriminative regions, including soft pixel and hard regional attention in HA-CNN.The passage describes attention as an automatically learned high-level mask.
  • GAN-based methods: GAN-based methods enrich ReID data or bridge domain and camera gaps by generating images.The related work includes GAN, PTGAN, and CamStyle.
  • Re-ranking methods: Re-ranking methods apply post-processing to retrieval results using affinity or Jaccard-distance information.The paper describes re-ranking as a critical step for improving retrieval accuracy.
  • Baseline framework: The standard baseline provides a widely used starting point, while the paper adds training tricks and BNNeck without changing most of the model architecture.The section states that the proposed BNNeck and center loss receive particular attention.

A. Warmup Learning Rate

The baseline uses warmup and scheduled learning-rate decay, alongside random erasing augmentation to address occlusion and improve generalization.

  • A. Warmup Learning Rate: 10 epochs of warmup linearly increase the learning rate from 3.5 × 10^-5 to 3.5 × 10^-4.The schedule then decays the rate at epochs 40 and 70.
  • A. Warmup Learning Rate: 3.5 × 10^-5 and 3.5 × 10^-6 are used after learning-rate decay at the 40th and 70th epochs, respectively.The schedule continues through epoch 120.
  • Training Tricks: Random erasing augmentation randomly masks a rectangle in training images with manually set probability p to address occlusion and improve generalization.The examples show original training images alongside their processed versions.
  • Training Pipeline: Fig. 2 presents the pipelines of the standard baseline and the modified baseline.The modified baseline includes the paper's training strategies and model architecture.

C. Label Smoothing

The paper describes ID-loss training and its overfitting challenge, then motivates BNNeck by separating classification and metric objectives into distinct feature spaces.

  • C. Label Smoothing: The ID loss is computed from image ID labels and class-wise prediction logits.The IDE network's last layer outputs ID prediction logits.
  • C. Label Smoothing: Person ReID testing identities do not appear during training, making prevention of overfitting to training IDs important.Label smoothing is introduced as a technique for classification-task overfitting.
  • E. BNNeck: The standard neck combines ID and triplet losses on the same feature f, despite their inconsistent embedding-space targets.ID loss is associated with cosine distance, whereas triplet loss uses Euclidean distance.
  • E. BNNeck: BNNeck adds a batch-normalization layer after the features and before the classifier, using pre-BN features for triplet loss and post-BN features for ID loss.This assigns the two losses to different feature spaces.
  • E. BNNeck: BNNeck uses the post-BN feature fi for inference, where cosine distance achieves better performance than Euclidean distance.The experiments report that BNNeck improves ReID performance by a large margin.

F. Center Loss

The paper identifies limitations of triplet-based ranking for enforcing intra-class compactness and introduces center loss to pull features toward class centers. Center loss is combined with ID and triplet losses in the model.

  • Task motivation: Ranking evaluation can overlook clustering quality: two cases may have identical rankings, while the more compact positive pairs are easier to separate with a tracking threshold.This motivates addressing intra-class compactness beyond retrieval metrics.
  • Triplet-loss limitation: Triplet loss can assign the same value to pairs with different absolute distances, so it ignores intra-class compactness.For dp = 0.3, dn = 0.5 and dp = 1.3, dn = 1.5, the triplet loss is 0.1 in both cases.
  • Center-loss formulation: Center loss simultaneously learns a center for each class and penalizes distances between features and their corresponding centers.It is introduced to compensate for drawbacks of triplet loss.
  • Center-loss effect: Minimizing center loss increases intra-class compactness by characterizing intra-class variations.For a mini-batch, y_j identifies the image label, c_yj its class center, and B the batch size.
  • Combined objective: The model combines ID loss, triplet loss, and center loss, with β weighting center loss at 0.0005 in the baseline.The combined objective is stated as L = L_ID + L_Triplet + βL_C.

IV. EXPERIMENT

Experiments evaluate the baseline and training tricks on Market1501 and DukeMTMC-reID, including standard and cross-domain settings. The combined tricks substantially improve performance, while BNNeck provides the largest individual boost and remains effective across domains except REA.

  • A. Datasets: The models are evaluated on Market1501 and DukeMTMC-reID using rank-1 accuracy and mAP.Both datasets are described as widely used and large scale.
  • A. Datasets: Market1501 contains 32,217 images from 1,501 identities across six camera views, with single-query evaluation.Its training set has 12,936 images and its testing set has 19,732 images.
  • A. Datasets: DukeMTMC-reID contains 36,411 images from 1,404 identities across eight camera views, also using single-query evaluation.The training and testing sets each contain images from 702 identities.
  • B. Influences of Each Trick: 87.7% and 79.7% rank-1 accuracy are achieved by Baseline-S on Market1501 and DukeMTMC-reID, respectively.Baseline-S represents the standard baseline.
  • B. Influences of Each Trick: 94.5% rank-1 and 85.9% mAP are achieved on Market1501, while DukeMTMC-reID reaches 86.4% rank-1 and 76.4% mAP after combining the tricks.The improvement exceeds 10% mAP over the standard baseline without increasing training time.
  • B. Influences of Each Trick: BNNeck provides a greater performance boost than the other individually added tricks, especially on DukeMTMC-reID.The evaluated tricks include warmup, REA, label smoothing, stride change, BNNeck, and center loss.
  • Cross-domain evaluation: Most tricks except REA are also effective for cross-domain ReID, whereas REA is inferred to hurt target-domain performance by masking source-domain image regions.The cross-domain evaluation uses Market1501-to-DukeMTMC-reID transfer among its reported settings.

C. Analysis of BNNeck

The ablations show that BNNeck is most effective when triplet loss acts before batch normalization and ID loss acts after it, separating their feature spaces. Visualizations indicate that the two losses have inconsistent distributions and that BN smooths ID features while harming triplet-feature compactness.

  • Different neck structures: BNNeck achieves the best performance on both benchmarks among the tested neck structures.Without batch normalization, combining ID and triplet losses is better than using either alone; the BN layer helps ID loss but not triplet loss.
  • BNNeck mechanism: BNNeck separates triplet and ID losses into two different feature spaces rather than forcing both losses to optimize the same feature.The neck designs and corresponding visualizations are presented in Figs. 8 and 9.
  • Feature distributions: ID and triplet losses produce different feature distributions, indicating that they are inconsistent in one embedding space.On MNIST, their distributions are described as affine and clustered rather than aligned.
  • Feature distributions: Integrating ID and triplet losses stretches the clustered distribution into a tadpole shape.The combined distribution is more distorted than the separate loss distributions.
  • BNNeck mechanism: The BN layer smooths or normalizes feature distributions and enhances ID-loss intra-class compactness while reducing triplet-loss compactness.This motivates placing triplet loss before BN and ID loss after BN in the proposed structure.

4) Two feature space of BNNeck:

The two BNNeck features behave differently under distance metrics because ID loss directly constrains the post-BN feature, while triplet loss favors the pre-BN feature. Center loss improves clustering but can reduce ranking accuracy.

  • Two feature spaces: Cosine distance performs better than Euclidean distance for ft, while fi is clearly separable through ID-loss hyperplanes.The authors select fi with cosine distance for inference.
  • Two feature spaces: BNNeck significantly improves person ReID performance, and fi with cosine distance is used for retrieval during inference.The selected feature is the one directly constrained by ID loss after batch normalization.
  • Center loss: Center loss reduces the intra-class-to-inter-class distance ratio as β increases, improving clustering effect on Market1501 and DukeMTMC-reID.R decreases from 0.407 to 0.311 on Market1501 and from 0.424 to 0.363 on DukeMTMC-reID as β rises from 0 to 0.5.
  • Center loss: At β = 0.5, center loss yields the best clustering effect but the worse rank-1 and mAP accuracies.This illustrates a trade-off between feature clustering and retrieval ranking performance.

E. Comparison to Other Baselines

The proposed baseline outperforms several established baselines and exceeds part-based or global-feature alternatives on both evaluated datasets. It also substantially boosts multiple reproduced state-of-the-art methods, showing that baseline quality affects reported method performance.

  • Comparison to state-of-the-art methods: 94.5% rank-1 accuracy and 85.9% mAP are achieved on Market1501 using the proposed strong baseline with global features.Pyramid reaches 92.8% rank-1 and 82.1% mAP with global features, while BFE combines two branches.
  • Experimental comparison: The reproduced comparisons use the same reported losses for each method and include author-reported values in parentheses.This setup provides a reference for separating baseline effects from method-specific effects.
  • Baseline meets state-of-the-art: The baseline boosts k-reciprocal re-ranking, PCB, AlignedReID++, and CamStyle by a large margin.The mAP gain for k-reciprocal re-ranking reaches +30.6% on Market1501.
  • Baseline meets state-of-the-art: The baseline can improve existing methods, although the gain is limited for MGN because its original accuracy is already too high to improve.CamStyle with the proposed baseline outperforms the reported CamStyle result but not Baseline1.

H. Performance of Different Backbones

Backbone choice materially affects performance, with deeper and larger backbones generally performing better, while attention and grouped convolution provide smaller gains. The baseline and BNNeck conclusions are evaluated primarily with ResNet50, and the paper notes unresolved cross-domain behavior.

  • Backbone performance: ResNet101 outperforms ResNet18 by 2.8% rank-1 and 9.3% mAP on Market1501.The experiments also evaluate ResNet, SeRes-Net, SeResNeXt, and IBNNet backbones.
  • Backbone performance: Channel attention in SeNet and group convolution in ResNeXt enhance performance by a slight margin.The paper contrasts these modest architectural gains with the larger effect of backbone depth and size.
  • Cross-domain performance: IBN-Net-a is suitable for same-domain tasks, whereas IBN-Net50-b is suitable for cross-domain tasks.IBN-Net50-a obtains 40.0% rank-1 and 25.1% mAP for M→D, and 52.9% rank-1 and 25.1% mAP for D→M.
  • Conclusions: The strong baseline reaches 94.5% rank-1 accuracy and 85.9% mAP on Market1501 using only an added BN layer for a standard baseline.The authors describe this as the best performance achieved by global features from a single backbone.
  • Outlook: The study leaves unresolved why REA reduces cross-domain performance in the authors’ baseline.The authors identify this confusion as a topic for further research.
Loading 1906.08332v2…