[Paper Review] 1st Place Solution to VisDA-2020: Bias Elimination for Domain Adaptive Pedestrian Re-identification

Jianyang Gu, Hao Luo+7 more|arXiv (Cornell University)|Dec 25, 2020

Video Surveillance and Tracking Methods14 references4 citations

TL;DR

This paper presents a two-stage domain adaptation framework for pedestrian re-identification that addresses inter-domain domain gap and intra-domain camera bias. By leveraging SPGAN-transferred data and CamStyle data augmentation, combined with a dual-domain training strategy using pseudo-labeling via DBSCAN and progressive refinement, the method achieves 76.56% mAP and 84.25% rank-1 on the VisDA-2020 test set, outperforming all competitors.

ABSTRACT

This paper presents our proposed methods for domain adaptive pedestrian re-identification (Re-ID) task in Visual Domain Adaptation Challenge (VisDA-2020). Considering the large gap between the source domain and target domain, we focused on solving two biases that influenced the performance on domain adaptive pedestrian Re-ID and proposed a two-stage training procedure. At the first stage, a baseline model is trained with images transferred from source domain to target domain and from single camera to multiple camera styles. Then we introduced a domain adaptation framework to train the model on source data and target data simultaneously. Different pseudo label generation strategies are adopted to continuously improve the discriminative ability of the model. Finally, with multiple models ensembled and additional post processing approaches adopted, our methods achieve 76.56% mAP and 84.25% rank-1 on the test set. Codes are available at https://github.com/vimar-gu/Bias-Eliminate-DA-ReID

Motivation & Objective

• To address the large domain gap between synthetic source domain and real-world target domain in pedestrian Re-ID.
• To mitigate intra-domain biases caused by variations in camera viewpoints, illumination, resolution, and occlusion.
• To improve model generalization and discriminative ability under domain shift via data augmentation and domain adaptation.
• To achieve state-of-the-art performance on the VisDA-2020 benchmark for domain adaptive pedestrian Re-ID.

Proposed method

• Utilizes SPGAN to transfer synthetic source images to match the appearance of real target domain images, reducing inter-domain domain gap.
• Applies StarGAN-based data augmentation to generate diverse camera-style images (CamStyle), improving robustness to camera-specific variations.
• Employs a two-stage domain adaptation framework with shared backbone and domain-specific classifiers, training on both source and target domains simultaneously.
• Uses DBSCAN clustering to generate pseudo-labels for target domain data, with progressive refinement by adding high-confidence, low-sample classes in the second stage.
• Applies a two-stage pseudo-labeling strategy: first selecting top 500 classes by sample count, then adding 200 additional one-sample classes to improve discriminative learning.
• Integrates model ensembling across multiple backbones (ResNet50-ibn-a, ResNet101-ibn-a, HRNetv2-w18) and post-processing with weighted distance matrix fusion to boost final performance.

Experimental results

Research questions

• RQ1 How can the domain gap between synthetic source and real-world target domains be effectively reduced in pedestrian Re-ID?
• RQ2 To what extent does camera-style variation across surveillance cameras degrade model generalization, and how can it be mitigated?
• RQ3 Can progressive pseudo-labeling with DBSCAN improve model performance in unsupervised domain adaptation for Re-ID?
• RQ4 What is the contribution of data augmentation (SPGAN and CamStyle) versus domain adaptation in reducing domain shift?
• RQ5 How does model ensembling and post-processing enhance final Re-ID performance on the VisDA-2020 benchmark?

Key findings

• The proposed method achieved 76.56% mAP and 84.25% rank-1 on the VisDA-2020 test set, securing first place in the challenge.
• The ablation study showed that adding CamStyle data increased mAP from 24.7% (with SPGAN) to 30.7%, highlighting the importance of camera bias mitigation.
• Domain adaptation with two-stage pseudo-labeling boosted mAP from 44.9% to 48.6%, demonstrating the effectiveness of progressive refinement.
• Post-processing with weighted distance matrix fusion increased mAP by 22.3 percentage points compared to the baseline single model.
• Model ensembling across multiple backbones and image sizes improved mAP by over 5.5 percentage points compared to the best single model.
• The final model achieved the highest rank-1 accuracy (84.25%) among all competitors, confirming the robustness of the framework.