Aditya Kasliwal (Manipal Institute of Technology); Pratinav Seth (Manipal Institute of Technology); Sriya Rallabandi (Manipal Institute of Technology) *; Sanchit Singhal (Manipal institute of technology) *
- ~ Equal Contribution
Thermal imaging has numerous advantages over regular visible-range imaging since it performs well in low-light circumstances. Super-Resolution approaches can broaden their usefulness by replicating accurate high-resolution thermal pictures using measurements from low-cost, lowresolution thermal sensors. Because of the spectral range mismatch between the images, Guided Super-Resolution of thermal images utilizing visible range images is difficult. However, In case of failure to capture Visible Range Images can prevent the operations of applications in critical areas.
We present a novel data fusion framework and regularization technique for Guided Super Resolution of Thermal images. The proposed architecture is computationally inexpensive and lightweight with the ability to maintain performance despite missing one of the modalities, i.e., highresolution RGB image or the lower-resolution thermal image, and is designed to be robust in the presence of missing data. The proposed method presents a promising solution to the frequently occurring problem of missing modalities in a real-world scenario.
Official implementation of https://arxiv.org/abs/2304.01243 paper.
$ git clone https://github.com/Kasliwal17/ThermalSuperResolution.git
$ cd ThermalSuperResolution
- Pytorch 1.11.0
- Segmentation-models-pytorch
- wandb
$ python -m src.train