Object counting models suffer when deployed across domains with differing density variety, since density shifts are inherently task-relevant and violate standard domain adaptation assumptions.To address this, we propose a theoretical framework of conditional feature alignment.We first formalize the notion of conditional divergence by partitioning each domain into subsets (e.g., object vs. background) and measuring divergences per condition.We then derive a joint error bound showing that, under discrete label spaces treated as condition sets, aligning distributions conditionally leads to tighter bounds on the combined source-target decision error than unconditional alignment.These insights motivate a general conditional adaptation principle: by preserving taskrelevant variations while filtering out nuisance shifts, one can achieve superior cross-domain generalization for counting.We provide both defining conditional divergence then proving its benefit in lowering joint error and a practical adaptation strategy that preserves task-relevant information in unsupervised domainadaptive counting.We demonstrate the effectiveness of our approach through extensive experiments on multiple counting datasets with varying density distributions.The results show that our method outperforms existing unsupervised domain adaptation methods, empirically validating the theoretical insights on conditional feature alignment.
This paper addresses the challenge of unsupervised domain-adaptive object counting by developing a theoretical framework focused on conditional feature alignment. The authors introduce the concept of conditional divergence, which partitions data into object-specific and background-specific subsets, allowing for a more accurate measure of divergence relevant to the counting task. They derive joint error bounds that demonstrate how conditional alignment leads to tighter guarantees compared to unconditional alignment, particularly important in scenarios where object density is task-relevant. The paper proposes a condition-driven alignment method that uses pseudo-label maps to segment images and adheres to task-relevant density variations while filtering out irrelevant shifts. The results from extensive experiments on various counting benchmarks indicate that this approach consistently outperforms existing state-of-the-art methods, validating the theoretical insights proposed. The contributions of this work are summarized as follows: theoretical framework of conditional divergence, a condition-driven alignment method that maintains task relevancy, a consistency mechanism to refine predictions without annotations, and extensive empirical validation showing superior counting performance across multiple scenarios.
This paper employs the following methods:
- Condition-driven alignment
- Pseudo-label maps
- Conditional divergence
The following datasets were used in this research:
- JHUCrowd++
- ShanghaiTech
- VGG
- ADI
- DCC
- Proposed method outperformed existing unsupervised domain adaptation methods
- Lower MAE and RMSE compared to state-of-the-art methods
The authors identified the following limitations:
- Existing domain adaptation methods fail to account for task-relevant density variations
- Need for enhanced pseudo-partition generation and wider applicability of conditional divergence framework
- Number of GPUs: 1
- GPU Type: NVIDIA RTX 3090
- Compute Requirements: epoch settings with a learning rate of 1e-6, weight decay of 1e-4, and training scalar adjustments for different datasets