Research map / 3D vision / Gaussian splatting
NTR-Gaussian: Nighttime Dynamic Thermal Reconstruction with 4D Gaussian Splatting Based on Thermodynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NTR-Gaussian, a two-stage 4D Gaussian Splatting method that predicts thermodynamic parameters (emissivity, convective heat transfer coefficient, heat capacity) with networks and integrates a heat-balance ODE to reconstruct nighttime dynamic thermal temperature fields, and releases the NTR aerial TIR dataset.
Problem
Static 3D thermal reconstruction methods capture only a single time period and cannot predict or analyze how scene surface temperature changes over the night.
Approach
Combine 4D Gaussian Splatting with a thermodynamic model: a Temp Net gives per-Gaussian temperature over time, and a Thermal Net predicts emissivity, convective coefficient and heat capacity that drive a differential heat equation whose integral is enforced to match the rendered temperature.
Key results
- NTR-Gaussian achieves average PSNR 34.61, SSIM 0.964 and MAE 1.065 across the four scenes, outperforming 3DGS, 4DGS and Thermal3D-GS.
- Predicted temperature error is within 1 degree Celsius according to the paper.
- Ablation on S1 shows removing the smoothing loss, spatial semantic features, or two-stage training lowers PSNR and raises MAE (e.g., full model MAE 0.947 vs 1.270 without smoothing loss).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use thermodynamics-based dynamic temperature prediction | NTR-Gaussian · S1 | MAE: 1.106 -> 0.947 | Helped |
| Use thermodynamics-based dynamic temperature prediction | NTR-Gaussian · S2 | MAE: 1.766 -> 1.709 | Helped |
| Use thermodynamics-based dynamic temperature prediction | NTR-Gaussian · S3 | MAE: 1.276 -> 0.959 | Helped |
| Use thermodynamics-based dynamic temperature prediction | NTR-Gaussian · S4 | MAE: 1.432 -> 1.065 | Helped |
| Add smoothing loss (ablation) | NTR-Gaussian · S1 | MAE: 1.270 -> 0.947 | Helped |
| Add spatial semantic features (ablation) | NTR-Gaussian · S1 | MAE: 1.498 -> 0.947 | Helped |
| Use two-stage training instead of single-stage training (ablation) | NTR-Gaussian · S1 | MAE: 1.496 -> 0.947 | Helped |
Surprising findings
- Training takes about 33402 s and 334 GB memory, far more than baselines (e.g., 3DGS 333 s, 8.4 GB).
Limitations
- Requires individualized optimization per scene and generalizes poorly across scenes and seasons.
- Does not model environmental factors such as humidity and wind speed.
- High training time and memory cost.
Experimental setup
Datasets: NTR dataset
Benchmarks: NTR dataset
Metrics: PSNR SSIM MAE (temperature) training time memory FPS
Baselines: 3DGS 4DGS Thermal3D-GS
Builds on: 3D Gaussian Splatting 4D Gaussian Splatting Feature 3DGS
Closest papers by meaning
- ThermalGaussian: Thermal 3D Gaussian Splatting (ICLR 2025 · similarity 0.87)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.83)
- A Theory of Joint Light and Heat Transport for Lambertian Scenes (CVPR 2024 · similarity 0.83)
- Gaussian Splatting for Efficient Satellite Image Photogrammetry (CVPR 2025 · similarity 0.83)
- DEAL: Data-Efficient Adversarial Learning for High-Quality Infrared Imaging (CVPR 2025 · similarity 0.82)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.