Core Concepts
ThermoNeRF proposes a novel approach using Neural Radiance Fields for rendering RGB and thermal views jointly, addressing challenges in thermal scene reconstruction.
Abstract
Introduction to ThermoNeRF and its application in thermal scene reconstruction.
Challenges with existing methods relying on RGB images for 3D geometry reconstruction.
Proposal of ThermoNeRF using paired RGB and thermal images for accurate synthesis.
Introduction of ThermoScenes dataset to address the lack of benchmark datasets.
Contributions of ThermoNeRF in achieving accurate thermal image synthesis.
Comparison with Nerfacto method showing significant improvement in temperature estimation.
Evaluation metrics including MAE, PSNR, SSIM for both thermal and RGB views.
Ablation study comparing different training strategies and highlighting the effectiveness of joint multimodal training in ThermoNeRF.
Stats
平均絶対誤差は1.5°Cで、Nerfactoと比較して50%以上の改善が見られた。
ThermoScenesデータセットには、10のシーンが含まれており、RGBとthermal画像のペアが提供されている。
テストセット全体でPSNRとSSIMの値が向上し、ThermoNeRFは他の手法よりも優れた性能を示した。