Source-linked AI summary
Bag of Tricks and A Strong Baseline for Deep Person Re-identification
Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, Wei Jiang
TL;DR
Person ReID research often relies on complex network structures and multi-branch features, while effective training tricks are scattered across papers and source code. This paper collects and evaluates those tricks, combines six into a strong baseline, and achieves 94.5% rank-1 accuracy and 85.9% mAP on Market1501 using global features.
Problem
Effective training tricks for person ReID are scattered across papers and open-source projects, while many state-of-the-art methods use complex network structures and multi-branch features.
Method
The paper collects and evaluates training tricks, selects six, designs BNNeck, and incorporates the tricks into a widely used ReID baseline.
Results
94.5% rank-1 accuracy and 85.9% mAP are achieved on Market1501 using only global features from a ResNet50 backbone.
Takeaways & Limitations
The authors suggest that training tricks can speed ReID research exploration, while noting that many effective tricks remain undiscovered.
Takeaways & Limitations
The work does not aim merely to improve performance, and the authors state that many effective training tricks remain undiscovered.
Abstract
from arXiv · showhide
This paper explores a simple and efficient baseline for person re-identification (ReID). Person re-identification (ReID) with deep neural networks has made progress and achieved high performance in recent years. However, many state-of-the-arts methods design complex network structure and concatenate multi-branch features. In the literature, some effective training tricks are briefly appeared in several papers or source codes. This paper will collect and evaluate these effective training tricks in person ReID. By combining these tricks together, the model achieves 94.5% rank-1 and 85.9% mAP on Market1501 with only using global features. Our codes and models are available at https://github.com/michuanhaohao/reid-strong-baseline.
1. Introduction
The paper argues that strong person ReID performance can come from collecting and evaluating effective training tricks rather than relying on complex architectures or multi-branch features. It combines selected tricks into a global-feature baseline and examines their implications for fair comparisons and practical deployment.
- Motivation: Deep ReID methods often use complex network structures and concatenate multi-branch features, while effective training tricks are scattered across papers and source code.The paper collects and evaluates these tricks systematically.
- Motivation: Most surveyed CVPR2018 and ECCV2018 baselines were built on weak baselines: only two of 23 Market1501 baselines exceeded 90% rank-1 accuracy.On DukeMTMC-reID, no surveyed baseline exceeded 80% rank-1 accuracy or 65% mAP.
- Motivation: Training tricks can account for improvements attributed to methods, so reviewers should consider them when comparing academic ReID results.The paper states that understated tricks can make method effectiveness appear exaggerated.
- Practical scope: The paper targets simple, effective models because local-feature concatenation and semantic modules add consumption and can reduce retrieval speed.Its stated industry goal is to improve models without too much extra consumption.
- Approach: The paper combines six selected tricks with a widely used baseline and evaluates individual improvements on two widely used ReID datasets.It also explores the effects of image size and batch size on model performance.
- Result: 94.5% rank-1 accuracy and 85.9% mAP are achieved on Market1501 using a ResNet50 baseline with only global features.The paper presents this as a strong baseline intended to support academia and industry.
2. Standard Baseline
The standard baseline uses a ResNet50 backbone with identity classification and metric-learning objectives. Its training pipeline specifies identity-balanced sampling, image preprocessing, feature and logit outputs, and Adam optimization.
- Architecture: The standard baseline uses a ResNet50 backbone initialized with ImageNet-pretrained parameters and an identity-classification layer sized to the training identities.N denotes the number of identities in the training dataset.
- Sampling: Each training batch randomly samples P identities and K images per identity, with P = 16, K = 4, and B = P×K.The resulting batch contains 64 images.
- Preprocessing: Images are resized to 256 × 128 pixels, padded by 10 pixels, randomly cropped, and horizontally flipped with probability 0.5.The pipeline then converts pixels to [0, 1] and applies channel-wise RGB normalization.
- Objectives: The model outputs ReID features f and identity-prediction logits p.The features support triplet loss, while the logits support cross-entropy loss.
- Optimization: Adam optimization starts at learning rate 0.00035, decreases by 0.1 at epochs 40 and 70, and runs for 120 epochs.The triplet-loss margin is set to 0.3.
3. Training Tricks
The paper extends a standard ReID baseline with training strategies that address optimization, augmentation, classification, feature resolution, embedding geometry, and intra-class compactness.
- Overview: The modified baseline adds warmup, random erasing, label smoothing, stride change, BNNeck, and center loss without changing most of the baseline architecture.These tricks are introduced as training strategies and model refinements for the standard baseline.
- Optimization and augmentation: Warmup linearly increases the learning rate from 3.5 × 10^-5 to 3.5 × 10^-4 during the first 10 epochs before scheduled decays.The rate is then reduced at the 40th and 70th epochs.
- Optimization and augmentation: Random Erasing Augmentation randomly selects and overwrites a rectangular image region to address occlusion and improve model generalization.The study uses erasing probability 0.5, area ratio 0.02 < Se < 0.4, and aspect-ratio bounds r1 = 0.3 and r2 = 3.33.
- Classification and feature learning: Label smoothing replaces hard identity targets with softened targets, using ε = 0.1 to make the model less confident on training identities.The paper motivates it as a way to prevent overfitting in the one-shot ReID setting.
- Classification and feature learning: Changing ResNet50’s last stride from 2 to 1 enlarges the feature map from 8 × 4 to 16 × 8 with light computation and no extra training parameters.The paper states that the higher spatial resolution brings significant improvement.
- Embedding geometry: BNNeck inserts batch normalization between features and classifier layers, using pre-normalization features for triplet loss and normalized features for ID loss.The design separates the feature roles of the two losses and uses the normalized feature with cosine distance during inference.
- Embedding geometry: Center loss learns a center for each class and penalizes feature-to-center distances, increasing intra-class compactness alongside the other losses.Its balanced weight β is set to 0.0005 in the experiments.
4. Experimental Results
Experiments evaluate progressively added training tricks on Market1501 and DukeMTMC-reID, including BNNeck ablations, cross-domain transfer, and comparisons with state-of-the-art methods.
- Same-domain evaluation: 94.5% rank-1 accuracy and 85.9% mAP are achieved on Market1501 after adding six training tricks to the standard baseline.On DukeMTMC-reID, the resulting model reaches 86.4% rank-1 accuracy and 76.4% mAP.
- Same-domain evaluation: BNNeck produces the largest performance boost among the evaluated tricks, especially on DukeMTMC-reID.The combined tricks improve the standard baseline by more than 10% mAP in other works, while adding only an extra BN layer without increasing training time.
- BNNeck analysis: BNNeck features are evaluated with Euclidean and cosine distances; cosine distance performs better for fi, while both metrics are similar for ft.The model uses fi with cosine distance for retrieval during inference.
- Cross-domain evaluation: Cross-domain experiments find that warmup, label smoothing, and BNNeck significantly improve transfer, whereas stride change and center loss have little impact.Random Erasing Augmentation harms cross-domain performance; without REA, the modified baseline reaches 41.4% and 54.3% rank-1 accuracy on the two datasets.
- State-of-the-art comparison: Using only global features, the strong baseline reaches 94.5% rank-1 accuracy and 85.9% mAP on Market1501 and outperforms AWTL by more than 10% mAP on both datasets.With k-reciprocal re-ranking, it reaches 94.1% mAP on Market1501 and 89.1% mAP on DukeMTMC-reID.
5. Supplementary Experiments
Supplementary experiments examine batch composition and image size as additional training factors, finding no specific batch-size conclusion and little image-size effect.
- Influences of the Number of Batch Size: Batch size is defined as B = P × K, where P is the number of identities and K is the number of images per identity.
- Influences of the Number of Batch Size: The batch-size experiments yield no specific conclusion, although a slight trend suggests larger batches may benefit performance.The authors infer that larger K helps mine hard positive pairs, while larger P helps mine hard negative pairs.
- Influences of Image Size: Four tested image sizes produce similar performance on both datasets when P = 16 and K = 4.The authors consider image size not an important factor for ReID model performance.
6. Conclusions and Outlooks
The paper collects and evaluates effective training tricks to build a strong person ReID baseline using global features. It reports strong Market1501 performance while identifying further experiments needed to understand the tricks and extend the baseline.
- The authors collect effective training tricks and design a strong baseline for person ReID.
- Experiments evaluate each trick's influence on both same-domain and cross-domain ReID tasks.
- 94.5% rank-1 accuracy and 85.9% mAP are achieved on Market1501 using only global features.
- The authors state that many effective training tricks remain undiscovered and invite researchers to share and evaluate additional tricks.
- Future work will analyze trick principles, test replacing BNNeck with L2 normalization, examine extensions to PCB, MGN, and AlignedReID, and add visualization.