Thermal homography estimation supports key inspection tasks in photovoltaic systems, including temporal fault tracking, aerial mosaicking, and multi modal image fusion. Thermal images contain uniform temperature patterns with low contrast and limited texture. These properties differ from RGB imagery that dominates existing homography estimation benchmarks. This paper introduces the first systematic benchmark of homography estimation methods (15 feature-based and 7 deep-learning-based methods) with 6 metrics on thermal photovoltaic images. The benchmark also evaluates 29 colormaps and provides a complete comparison under controlled settings. The results show a performance gap between detector-free vs detector-based models, and conventional vs deep learning methods. Detector free models achieve strong performance with success rate up to 97.09% by removing the dependence on corner structure. Deep learning methods surpass traditional approaches by 56.2% in success rate. Colormap selection produces large performance variations for traditional methods, while deep learning models maintain more consistent behavior and show reduced sensitivity to visual changes. Nevertheless, traditional pipelines remain valuable in GPU constrained and resource-limited environments. Altogether, the benchmark framework advances thermal image analysis for photovoltaic applications and provides a foundation for reliable deployment in industrial inspection.