<![CDATA[Computer vision provides a promising solution to evaluate housing conditions. However, the differences between indoor and outdoor visual environments may cause domain gaps that affect the generalization ability of the model. This study explores the effect of domain gaps on the detection of housing physical conditions using You Only Look Once version 11 (YOLOv11) by proposing an integrated evaluation framework that combines cross-domain evaluation, EigenCAM-based model interpretability, and illumination variation analysis. A total of 2,712 housing images were classified into Indoor, Outdoor and Mixed datasets and evaluated under uniform experimental conditions. From the results, the Indoor model obtained a 42.77% relative mAP50 drop when evaluated on the Outdoor dataset, thus proving the existence of a significant domain gap. The Mixed model achieved the most consistent performance across all evaluation scenarios, although this improvement may also be partially influenced by the larger training dataset used in the Mixed configuration. EigenCAM analysis showed that single-domain models were more reliant on contextual visual cues while the Mixed model was consistently more attuned to relevant housing elements. The illumination analysis showed that the differences in brightness between the domains were not large, suggesting that context bias is more likely than illumination variation to explain the observed performance degradation. These results demonstrate that the combination of cross-domain evaluation with model interpretability provides a more complete understanding of domain gap effects and enables the creation of more robust computer vision models for housing condition assessment.]]>
