On the Effect of Domain-Adversarial Supervision Placement for Cross-Sensor 6-D Pose Estimation

Domain-adversarial training is a common approach to mitigate domain shift in deep learning, yet architectural design choices such as the placement of adversarial supervision are often treated heuristically. We present a systematic empirical assessment of gradient reversal-based domain-adversarial training applied to cross-sensor 6D pose estimation, focusing on the placement of domain classification heads within a multi-modal red, green, and blue(RGB)–point cloud fusion network. Using a simplified bidirectional RGB–point cloud fusion network, we evaluate domain head placement at multiple depths, including modality-specific encoders and intermediate fusion stages, under a strict multi-source training protocol with no access to target-domain data. Experiments are conducted on controlled synthetic datasets with multiple runs per configuration and complemented by real-world cross-sensor RGB-D evaluations. Across all settings, performance differences between domain head placements are dominated by run-to-run variability, with no statistically significant advantage for any particular placement. Real-world experiments exhibit similarly small differences and qualitatively align with synthetic results. These findings indicate that, for the studied architecture, gradient reversal-based domain-adversarial training is largely insensitive to the precise placement of the domain head, suggesting that both early and late integration constitute viable design choices. This robustness provides practical flexibility in network design, supports reproducible evaluation across placements, and motivates future work on complementary adaptation mechanisms that can be combined with adversarial supervision.