Source-linked AI summary
Cross-Dataset Person Re-Identification via Unsupervised Pose Disentanglement and Adaptation
Yu-Jhe Li, Ci-Siang Lin, Yan-Bo Lin, Yu-Chiang Frank Wang
TL;DR
Cross-dataset re-ID seeks to generalize from labeled source images to an unlabeled target domain, while prior approaches have difficulty preserving pose across datasets. PDA-Net jointly learns domain-invariant and pose-disentangled representations without target identity or pose supervision, and experiments on two benchmark datasets show remarkable improvements over existing works.
Problem
Cross-dataset re-ID must transfer discriminative information from a labeled source domain to an unlabeled target domain, but existing approaches struggle to preserve pose information across datasets.
Method
PDA-Net jointly learns domain-invariant image features and disentangled pose features, using image and pose inputs to support pose-guided recovery and cross-domain translation without target labels.
Results
Experiments on two benchmark datasets show remarkable improvements over existing cross-dataset re-ID works, with quantitative and qualitative results confirming favorable performance.
Takeaways & Limitations
PDA-Net provides a pose-guided, dataset-invariant representation for applying re-ID across domains without identity supervision in the target domain.
Takeaways & Limitations
The method does not use adversarial learning to derive domain-invariant features because it might instead produce pose-invariant features that perform poorly in cross-dataset re-ID.
Abstract
from arXiv · showhide
Person re-identification (re-ID) aims at recognizing the same person from images taken across different cameras. To address this challenging task, existing re-ID models typically rely on a large amount of labeled training data, which is not practical for real-world applications. To alleviate this limitation, researchers now targets at cross-dataset re-ID which focuses on generalizing the discriminative ability to the unlabeled target domain when given a labeled source domain dataset. To achieve this goal, our proposed Pose Disentanglement and Adaptation Network (PDA-Net) aims at learning deep image representation with pose and domain information properly disentangled. With the learned cross-domain pose invariant feature space, our proposed PDA-Net is able to perform pose disentanglement across domains without supervision in identities, and the resulting features can be applied to cross-dataset re-ID. Both of our qualitative and quantitative results on two benchmark datasets confirm the effectiveness of our approach and its superiority over the state-of-the-art cross-dataset Re-ID approaches.
1. Introduction
Cross-dataset re-ID addresses the need to generalize from labeled source images to an unlabeled target domain, where existing methods struggle to preserve pose information. PDA-Net jointly learns pose-disentangled, domain-invariant representations without target identity or pose labels, and performs favorably on two unsupervised cross-dataset tasks.
- Motivation: Person re-ID matches images of the same identity across nonoverlapping camera views, but conventional approaches require large labeled datasets.Cross-dataset re-ID instead transfers information from a fully labeled source domain to an unlabeled target domain.
- Proposed approach: PDA-Net jointly performs pose disentanglement and adaptation rather than sequential style transfer followed by feature extraction, improving image representation.The joint design is intended to alleviate limitations on the image variants observed by prior approaches.
- Limitations of prior work: Existing cross-dataset methods adapt style information across datasets, making pose information difficult to describe or preserve.Examples include image translation with CycleGAN or StarGAN and feature alignment with MMD.
- Proposed approach: PDA-Net learns domain-invariant features while disentangling pose information without target-domain identity or pose labels.Source labels are used to enforce re-ID performance, while pose information from either domain supports the unified framework.
- Results: Experiments on two challenging unsupervised cross-dataset re-ID tasks confirm favorable quantitative and qualitative performance against state-of-the-art approaches.The introduction reports the method as practically preferable based on improved performance over prior unsupervised cross-dataset methods.
2. Related Works
Related work spans supervised re-ID, cross-dataset adaptation, and pose-guided methods. PDA-Net combines domain-invariant content representation with pose encoding and domain-specific image generation for cross-dataset re-ID.
- Supervised Person Re-ID: Supervised re-ID methods address viewpoint, pose, background clutter, and occlusion, but do not easily transfer across datasets because visual differences remain unsuppressed.Representative approaches use pose transfer, CRF-based similarity metrics, and attention mechanisms.
- Cross-dataset Person Re-ID: Cross-dataset re-ID methods use hand-crafted features or domain adaptation, but differing identities, viewpoints, poses, clutter, and absent target labels can limit gains.Prior approaches include style translation, camera-invariance learning, and mid-level feature alignment.
- Pose-Guided Re-ID: Pose-guided re-ID methods improve pose representation, yet existing cross-dataset approaches may require prior pose knowledge or fail to encode pose in resulting features.Some pose-generation methods also rely on complex multi-branch designs that are not end-to-end trainable.
- PDA-Net: PDA-Net encodes domain-invariant content and pose features, then uses source- and target-specific generators for recovery and cross-domain translation.Domain discriminators preserve perceptual quality, while a pose discriminator supports pose disentanglement.
3. Proposed Method
PDA-Net learns domain-invariant person representations while disentangling pose across labeled source and unlabeled target domains. It combines pose encoding, domain-specific image generation, reconstruction, translation, distribution alignment, and source-label supervision for cross-dataset re-ID.
- Network architecture: PDA-Net encodes images into domain-invariant content features and pose maps into separate pose features for re-ID and pose disentanglement.The content encoder produces v_c, while the pose encoder produces v_p.
- Image generation: Domain-specific generators reconstruct or synthesize source- and target-domain images from content–pose feature pairs, including cross-domain combinations.The generators G_S and G_T receive content and pose features and output images in their associated domains.
- Inference and objective: At test time, the content encoder produces a domain- and pose-invariant query representation matched to gallery images using nearest-neighbor search in Euclidean distance.The training objective is cross-dataset re-ID without target-domain identity labels, using pose-guided cross-domain image translation and joint encoder–decoder training.
- Domain-invariant representation: MMD encourages similar content-feature distributions across source and target domains, while source triplet supervision separates identities in the content space.The triplet loss uses same-identity positives and different-identity negatives with margin m > 0.
- Pose-guided recovery: Pose-guided recovery uses source images with different poses but the same identity to preserve shared content while reconstructing a desired pose.The reconstruction terms use the L1 norm, which the paper states preserves image sharpness.
- Adversarial training: Source and target discriminators enforce perceptually realistic outputs in their respective domains during reconstruction and translation.The source discriminator distinguishes real source images from reconstructed and translated source-domain images, with an analogous target-domain discriminator.
4. Experiments
Experiments on Market-1501 and DukeMTMC-reID evaluate PDA-Net quantitatively and qualitatively, showing favorable cross-dataset re-ID performance and pose-guided image synthesis. Ablations associate performance gains with reconstruction, pose, domain, MMD, and shared pose-discriminator components.
- Quantitative Comparisons: Market-1501 evaluation reports Rank-1 accuracy=75.2% and mAP=52.6% in the single query setting.The comparison includes unsupervised and cross-dataset re-ID baselines.
- Quantitative Comparisons: DukeMTMC-reID evaluation reports Rank-1 accuracy=63.2% and mAP=45.1% in the single query setting.The results are reported as favorable against baseline and state-of-the-art unsupervised or cross-domain methods.
- Quantitative Comparisons: PDA-Net exceeds TAUDL by 11.5% in Rank-1 accuracy and 11.4% in mAP on Market-1501, without additional spatial or temporal information.On DukeMTMC-reID, it exceeds the second-best method by 1.5% in Rank-1 accuracy and 1.6% in mAP.
- Ablation Studies: The full model improves Rank-1 performance by roughly 20–25% on both benchmark datasets relative to the evaluated baseline configurations.The study attributes the gain to PDA-Net’s design for deriving domain-invariant and pose-invariant representations.
- Ablation Studies: Removing reconstruction loss causes 23% and 20% drops on Market-1501 and DukeMTMC-reID, respectively.The paper links these drops to the absence of explicit supervision for generating human-perceivable images and resulting image-level information loss.
- Ablation Studies: Removing pose loss causes about 20% re-ID performance drops in both settings, while disabling domain loss causes a 10% drop.A shared pose discriminator is preferred because pose guidance can be provided by both domains; visualizations also show within- and cross-domain pose-guided synthesis.
5. Conclusions
The paper concludes that PDA-Net jointly learns domain-invariant and pose-disentangled representations for cross-dataset re-ID. Experiments support improved performance and qualitative cross-domain image translation with pose manipulation.
- Conclusions: PDA-Net jointly learns domain-invariant and pose-disentangled visual representations for cross-dataset re-ID.The model uses image inputs from either domain and desirable pose information for pose-guided recovery or cross-domain translation.
- Conclusions: Experiments on two benchmark datasets show remarkable improvements over existing works and support PDA-Net for cross-dataset re-ID.Qualitative results confirm cross-domain image translation with pose properly disentangled and manipulated.