1. Qualitative Result
Ground Truth. Reference thermal sequences for the scenes shown above.
2. Dataset
Due to the limited availability of large-scale dynamic thermal datasets, we conduct our experiments using a combination of publicly available benchmarks, including NTR-Gaussian and TI-NSD. NTR-Gaussian is designed for nighttime dynamic thermal reconstruction and contains aerial thermal infrared (TIR) imagery from four outdoor scenes (two urban, two suburban), captured at multiple time intervals using a DJI Matrice 300 RTK equipped with a DJI H20T thermal camera. The data is collected at an altitude of 250 meters with a resolution of 640 × 512 pixels and covers diverse environments such as buildings, roads, farmlands, and water bodies. The dataset provides calibrated camera poses and paired synthetic RGB images, enabling multi view and time varying thermal scene modeling. While TI-NSD further complements this data as a larger benchmark, comprising 20 real thermal infrared video scenes spanning indoor, outdoor, and UAV scenarios for evaluating thermal novel view reconstruction techniques.
3. Quantitive Results
4. Future Directions
We plan to expand our evaluation beyond visual quality by introducing targeted experiments that test whether predicted temperature evolution is consistent with observed thermal trends in real scenes. This includes validating relative temperature changes on fixed surfaces across repeated UAV passes, measuring temporal consistency under partial observations, and comparing against simpler baselines to identify when dynamic modeling provides clear benefits. We also aim to explore lightweight validation signals and controlled scenarios that help assess physical plausibility without requiring dense ground truth temperature measurements. Together, these steps will clarify the practical scope of 4D thermal reconstruction and guide its use in applications where spatiotemporal modeling is most informative.