Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment

TL;DR

The paper tackles universal domain adaptive object detection (UniDAOD) under open-set and partial-set shifts, identifying that global-level domain-private alignment and cross-level probability heterogeneity hinder existing methods. It introduces Dual Probabilistic Alignment (DPA), comprising Global-level Domain Private Alignment (GDPA), Instance-level Domain Shared Alignment (IDSA), and Private Class Constraint (PCC), with domain probability modeled as a Gaussian to enable adaptive sampling and weighting. The authors provide a theoretical bound demonstrating the importance of domain-private alignment, and design GDPA to target global-private samples, IDSA to align domain-shared samples at the instance level, and PCC to mitigate negative transfer through cross-space centroid constraints. Extensive experiments on diverse benchmarks show that DPA achieves state-of-the-art performance across open-set, partial-set, and closed-set settings, highlighting its effectiveness and practical impact for robust UniDAOD in real-world scenarios.

Abstract

Domain Adaptive Object Detection (DAOD) transfers knowledge from a labeled source domain to an unannotated target domain under closed-set assumption. Universal DAOD (UniDAOD) extends DAOD to handle open-set, partial-set, and closed-set domain adaptation. In this paper, we first unveil two issues: domain-private category alignment is crucial for global-level features, and the domain probability heterogeneity of features across different levels. To address these issues, we propose a novel Dual Probabilistic Alignment (DPA) framework to model domain probability as Gaussian distribution, enabling the heterogeneity domain distribution sampling and measurement. The DPA consists of three tailored modules: the Global-level Domain Private Alignment (GDPA), the Instance-level Domain Shared Alignment (IDSA), and the Private Class Constraint (PCC). GDPA utilizes the global-level sampling to mine domain-private category samples and calculate alignment weight through a cumulative distribution function to address the global-level private category alignment. IDSA utilizes instance-level sampling to mine domain-shared category samples and calculates alignment weight through Gaussian distribution to conduct the domain-shared category domain alignment to address the feature heterogeneity. The PCC aggregates domain-private category centroids between feature and probability spaces to mitigate negative transfer. Extensive experiments demonstrate that our DPA outperforms state-of-the-art UniDAOD and DAOD methods across various datasets and scenarios, including open, partial, and closed sets. Codes are available at \url{https://github.com/zyfone/DPA}.

Figures (5)

• Figure 1: The visualization of domain probability in the domain discriminator. (a) As the number of domain-private categories increases, a more distinct gap emerges between the two domains at the global level, suggesting the alignment of the domain-private category. (b) With the increases of the domain-private category, the mean probability gap remains approximately constant at the instance level, indicating the alignment of domain-shared category.
• Figure 2: Illustration of the proposed DPA framework. (a) GDPA establishes the global-level embedding feature to sample the outlier in the feature space, then applies a CDF of Gauss distribution to weighting the probability distribution. (b) IDSA obtains the gradient norm of instance probability and models it as a Gauss distribution for sampling and weighting. (c) PCC obtains the domain-private common centroid and constraints distances of samples to the centroid between feature and probability spaces.
• Figure 3: Category-wise performance gain over the source-only model (classes are plane, bicycle, bird, boat, and bottle). Positive transfer is green, and negative transfer is red.
• Figure 4: Qualitative analysis of category alignment in terms of the mean probability gap: (a) global-level features and (b) instance-level features. The horizontal axis represents training iterations ($\times$100), and the vertical axis shows the probability of the domain discriminator. The benchmark is Pascal VOC to Clipart1k. ($\beta=25\%$).
• Figure 5: The weight quantitative analysis of global-level domain-private feature. The horizontal axis is training iteration ($\times$100), and the vertical axis is weight values $\frac{ \Phi_s/(1-\Phi_t)}{ \Phi_s+(1-\Phi_t)}$ in source and target domains.