Image Processing Group

This study evaluates deep-learning and Shape from Silhouette (SfS) methods for 3D reconstruction of smoke plumes. It demonstrates the deep-learning method's superiority in cases with limited camera views and calibration data, achieving high-quality reconstructions of semitransparent smoke without precise calibration. The research emphasizes the significance of preprocessing and data appearance for neural network efficacy. By improving 3D reconstruction techniques, this work aids in advancing wildfire tracking and environmental analysis, offering a practical approach for real-world applications in fire science.