Source-linked AI summary
Unsupervised Pre-training for Person Re-identification
Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, Dong Chen
TL;DR
Person Re-ID lacks large, diverse labeled datasets, and ImageNet pre-training leaves a domain gap. The paper introduces LUPerson and studies unsupervised pre-training with Re-ID-specific augmentation and contrastive-loss choices. The resulting generic representation improves supervised and unsupervised methods across four datasets, especially in small-scale or few-shot settings.
Problem
Existing person Re-ID datasets have limited scale and diversity because annotation is costly, while ImageNet pre-training has a large domain gap from Re-ID data.
Method
The paper builds LUPerson and studies unsupervised Re-ID pre-training through data augmentation strategies and temperature selection in contrastive learning.
Results
The pre-trained representation improves supervised and unsupervised Re-ID methods across CUHK03, Market1501, DukeMTMC, and MSMT17, with larger gains in small-scale or few-shot settings.
Takeaways & Limitations
Large-scale unsupervised pre-training can provide a generic Re-ID feature that benefits existing supervised and unsupervised Re-ID approaches.
Abstract
from arXiv · showhide
In this paper, we present a large scale unlabeled person re-identification (Re-ID) dataset "LUPerson" and make the first attempt of performing unsupervised pre-training for improving the generalization ability of the learned person Re-ID feature representation. This is to address the problem that all existing person Re-ID datasets are all of limited scale due to the costly effort required for data annotation. Previous research tries to leverage models pre-trained on ImageNet to mitigate the shortage of person Re-ID data but suffers from the large domain gap between ImageNet and person Re-ID data. LUPerson is an unlabeled dataset of 4M images of over 200K identities, which is 30X larger than the largest existing Re-ID dataset. It also covers a much diverse range of capturing environments (eg, camera settings, scenes, etc.). Based on this dataset, we systematically study the key factors for learning Re-ID features from two perspectives: data augmentation and contrastive loss. Unsupervised pre-training performed on this large-scale dataset effectively leads to a generic Re-ID feature that can benefit all existing person Re-ID methods. Using our pre-trained model in some basic frameworks, our methods achieve state-of-the-art results without bells and whistles on four widely used Re-ID datasets: CUHK03, Market1501, DukeMTMC, and MSMT17. Our results also show that the performance improvement is more significant on small-scale target datasets or under few-shot setting.
1. Introduction
Existing person Re-ID datasets are small and lack environmental diversity, while ImageNet pre-training leaves a domain gap. The paper introduces LUPerson and studies unsupervised pre-training choices that improve supervised and unsupervised Re-ID across datasets.
- Motivation: Existing Re-ID datasets are limited in images, identities, scenes, camera settings, and resolution, while ImageNet has a large domain gap from person Re-ID data.The largest cited datasets contain 126K images or 9,651 identities and fewer than 20 scenes.
- LUPerson: LUPerson contains 4M person images from over 200K identities and diverse environments, enabling the first large-scale unsupervised pre-training for person Re-ID.The images were extracted from 46K YouTube videos, and the dataset is 30× larger than MSMT17.
- Unsupervised pre-training: Color distortion harms Re-ID feature learning because color is a crucial identity clue, whereas RandomErasing benefits unsupervised training as in supervised Re-ID.The study also identifies contrastive-loss temperature as important for balancing discriminativity and hard-negative mining.
- Results: 3.5% mAP on Market1501, 2.7% on DukeMTMC, 2.0% on MSMT17, and 2.9% on CUHK03 are reported gains from the pre-trained model over strong supervised baselines.The improvements are reported using MGN on three datasets and BDB on CUHK03.
- Unsupervised pre-training: The paper makes generic unsupervised pre-training possible by investigating task-specific augmentation strategies and temperature selection in contrastive learning.This work uses those factors to study unsupervised Re-ID feature learning on LUPerson.
- Results: The unsupervised representation is general to supervised and unsupervised Re-ID methods and yields larger improvements with small-scale or few-shot training data.The paper reports consistent improvements on unsupervised methods based on the SpCL baseline.
2. Related Work
The related work discusses supervised Re-ID approaches and recent attempts to train unsupervised models directly on existing Re-ID datasets. It positions MoCo v2 pre-training as a starting point that requires Re-ID-specific investigation.
- Supervised Person Re-ID: Supervised Re-ID methods learn global, metric-learning, or part-based features, with MGN combining global classification and local triplet losses.The paper presents its pre-trained model as applicable to these representative supervised methods.
- Unsupervised Person Re-ID: Recent unsupervised Re-ID work trains models without labels on existing Re-ID datasets, while this paper uses MoCo v2 for unsupervised pre-training.Directly applying MoCo v2 to person Re-ID does not work well, motivating analysis of task-specific factors.
3. LUPerson: Large-scale Re-ID Dataset
LUPerson addresses the limited scale and diversity of existing person Re-ID datasets with a large unlabeled collection designed for generic feature learning. It contains over 4M images spanning diverse scenes, devices, lighting, ethnicities, and poses.
- Motivation: Existing Re-ID datasets contain fewer than 0.2M images, fewer than 10K identities, and fewer than 20 scenes with limited camera settings.These constraints make learning high-quality generic feature representations difficult.
- Dataset construction: LUPerson contains 4M images of more than 200K identities collected from 46K scenes, making it the largest person Re-ID dataset described.The dataset was built to approach ImageNet-scale data for person Re-ID.
- Dataset construction: The dataset was assembled from YouTube streetview videos gathered across the world’s top 100 cities and filtered to remove duplicates, short, static, and virtual-reality videos.The collection process retained 50,534 videos after filtering.
- Comparison with existing datasets: LUPerson is 30× larger than MSMT17 and includes static and dynamic camera views across streets, campuses, supermarkets, and sports fields.The dataset also uses varied recording devices, including video recorders, vlogs, smartphones, and surveillance cameras.
- Comparison with existing datasets: Its videos span morning, noon, and night and provide ethnic and pose diversity through collection across 100 cities and varied camera views.Example images illustrate variation in environment, scene, camera view, lighting, pose, race, and age.
4. Unsupervised Pre-training for Person Re-ID
The paper adapts unsupervised contrastive pre-training to person Re-ID, where generic MoCoV2 settings do not transfer well because the task depends on fine-grained identity cues. It redesigns augmentation and temperature choices, then calibrates feature use for downstream Re-ID methods.
- Motivation: Directly applying MoCoV2 does not learn a good person Re-ID representation, motivating task-specific redesign.Person Re-ID requires fine-grained details for robust person association.
- Contrastive Learning: Contrastive learning creates two augmented views of each image as a query-positive pair and treats views from other images as negatives.The loss uses temperature τ and K negative samples, with a queue maintaining a large dictionary.
- Data Augmentation: Disabling color jitter improves CUHK03 transfer by 0.6% mAP and 0.6% cmc1 because Re-ID relies heavily on color cues.Random grayscale and Gaussian blurring have limited impact in the reported investigation.
- Data Augmentation: Adding RandomErasing yields about a 0.8% mAP gain, with strength 0.6 optimal for unsupervised pre-training rather than 0.4 for existing Re-ID methods.The tables vary RandomErasing inclusion and maximum erased area.
- Temperature Tuning Strategy: Reducing temperature from τ = 0.2 to τ = 0.07 raises CUHK03 mAP from 71.1 to 74.7, whereas general image recognition declines from 66.2 to 62.9.The paper attributes the difference as a possible consequence of person Re-ID being more fine-grained.
- Transfer to Re-ID Methods: The resulting pre-trained representation benefits supervised and unsupervised Re-ID methods, with an extra batch-normalization layer used to recalibrate downstream features.This allows the same hyper-parameters as ImageNet-supervised counterparts in the described example.
5. Experiments
Experiments show that LUPerson unsupervised pre-training improves supervised and unsupervised Re-ID across multiple benchmarks, with especially large gains in small-data settings. Performance also increases with pre-training data scale before saturating.
- Improving Supervised Re-ID Methods: Unsupervised LUPerson pre-training improves three supervised Re-ID baselines by more than 4.2%, 3.5%, 2.7%, and 2% mAP on CUHK03, Market1501, DukeMTMC, and MSMT17, respectively.It outperforms supervised and unsupervised ImageNet pre-training in most cases.
- Comparison on Small-scale and Few-shot: LUPerson pre-training significantly boosts MGN when identities or images per identity are reduced, although few-shot gains become somewhat saturated as image counts decrease.The study evaluates small-scale and few-shot settings from 10% to 100% of the training data.
- Comparison on Small-scale and Few-shot: 64.6 mAP on 10% of Market1501 training data is 11.5 mAP higher than the ImageNet-supervised counterpart.This small-scale setting contains 1,170 images for 75 persons.
- Pre-training Data Scale: Learned representations strengthen as the LUPerson pre-training data scale increases, but performance tends to saturate once the scale is large enough.The authors suggest that greater network capacity would be needed to exploit more data further.
- Improving Unsupervised Re-ID Methods: LUPerson pre-training improves SpCL by 3.8 mAP on Market1501 and 2.2 mAP on DukeMTMC, while UDA gains are 2.8% for M→D and 1.2% for D→M mAP.These results establish a new state-of-the-art for USL in person Re-ID according to the paper.
- Comparison with State-of-the-Arts: Using a pre-trained ResNet50 model with MGN achieves state-of-the-art performance on Market1501, DukeMTMC, and MSMT17 without post-processing; BDB gains 2.9 mAP on CUHK03.A ResNet101 backbone yields better performance than ResNet50.
6. Conclusion
The paper releases LUPerson and uses it for unsupervised pre-training of general person Re-ID representations. Experiments show effectiveness across supervised and unsupervised approaches, with larger gains in small-scale or few-shot settings.
- 6. Conclusion: LUPerson addresses the limited scale and diversity of existing person Re-ID datasets with a large unlabeled dataset.The paper identifies video temporal information and end-to-end unsupervised Re-ID as directions for future work.
- 6. Conclusion: Unsupervised pre-training on LUPerson learns a general person Re-ID representation that benefits both supervised and unsupervised Re-ID approaches.The reported gains are bigger for small-scale or few-shot Re-ID datasets.