Abstract
Human parsing models often suffer from performance degradation when applied to new domains with different lighting, resolutions, or viewpoints. This paper presents a novel cross-domain human parsing framework that leverages adversarial learning to bridge the gap between a labeled source domain and an unlabeled target domain. By employing a dual-adversarial strategy—targeting both feature-level appearance and label-level structural priors—the proposed method achieves state-of-the-art performance without requiring manual annotations in the target domain. Experimental results on diverse datasets, including surveillance and indoor environments, demonstrate the effectiveness of the approach in handling significant domain shifts.
Introduction
Human parsing involves the fine-grained segmentation of human images into semantic parts such as limbs, clothing, and hair. While deep learning models have achieved high accuracy on standard benchmarks, their real-world utility is often limited by domain shift. For instance, a model trained on high-resolution fashion photography (source domain) typically fails when applied to low-quality surveillance footage or dimly lit indoor scenes (target domains).
The primary challenge lies in the variations in illumination, camera angles, and image quality. Traditionally, addressing this required expensive and time-consuming manual labeling for every new target environment. This work explores an unsupervised domain adaptation approach, aiming to transfer knowledge from a rich source dataset to unlabeled target domains.
Methods
The proposed framework utilizes a dual-adversarial approach to address domain shift at two distinct levels: visual appearance and structural logic.
1. Adversarial Feature Adaptation
To handle low-level appearance differences, the model incorporates a Feature Compensation Network ($C$). Unlike standard methods that seek a shared feature space, this network explicitly learns a "compensation" to map source features into the target space, accounting for environmental factors like blur or lighting. A Feature Adversarial Network ($A_f$) acts as a discriminator, attempting to distinguish between target features and compensated source features, driving the compensation network to produce domain-invariant representations.
2. Adversarial Structured Label Adaptation
While appearance varies, the underlying structure of the human body remains constant across domains. The Structured Label Adversarial Network ($A_l$) enforces these high-order spatial relationships. By learning "legal" configurations of human parts from the source ground truth, this discriminator forces the target domain's parsing predictions to adhere to realistic human structural priors (e.g., the head positioned above the torso).
The system is built on a VGG-16/DeepLab backbone and utilizes Least Squares GANs (LSGANs) to ensure stable training during the complex pixel-wise segmentation task.
Experimental Setup
The model was trained using the LIP (Look Into Person) dataset as the source domain, which contains over 50,000 images in unconstrained poses. Evaluation was conducted on several unlabeled target domains:
- Indoor Dataset: 1,900 images from a canteen with challenging lighting.
- Daily Video Dataset: 1,584 images from real-world scenes like roads and shops.
- PridA & PridB: Low-resolution surveillance camera trajectories.
Results and Discussion
The proposed "Feat.
- Lab. Adapt" model consistently outperformed "Source Only" baselines and traditional domain adaptation methods such as DANN.
- Quantitative Performance: On the Daily Video dataset, the model achieved an Average F1 score of 64.36%, notably surpassing a "Target Only" model trained with full supervision on that specific dataset (59.73%). This indicates that the framework effectively transfers the rich semantic knowledge of the large-scale source dataset while adapting to target-specific environmental conditions.
- Qualitative Robustness: The model demonstrated robustness in low-resolution and poorly lit scenarios, accurately identifying body parts where baseline models often failed due to background confusion. A key insight is that feature adaptation alone is insufficient; enforcing structural logic through label adaptation is critical for robust performance in the "wild."
Conclusion
This paper demonstrates that combining explicit feature compensation with structural label adaptation provides a powerful framework for unsupervised cross-domain human parsing. Future work may explore the integration of temporal context for video-based adaptation and the application of alternative GAN architectures, such as Wasserstein GANs, to further enhance training stability.
Summary updated