Datasets

VISWIR vQuasar relies on paired datasets of visible spectrum and SWIR images. This section explains the expected folder organization and provides guidelines for preparing your data.

Folder Structure

The dataset should be organized as follows:

  • data/visible/ : contains visible spectrum images (e.g. RGB).

  • data/swir/ : contains SWIR images (paired with visible).

  • data/ground_truth/ : optional ground truth annotations (for detection tasks).

Example layout:

data/
├── visible/
│   ├── img_001.png
│   ├── img_002.png
│   └── ...
├── swir/
│   ├── img_001.png
│   ├── img_002.png
│   └── ...
└── ground_truth/
    ├── img_001.xml
    ├── img_002.xml
    └── ...

Notes

  • Pairing: Visible and SWIR images must be paired and sorted consistently (e.g. img_001.png in both folders corresponds to the same scene).

  • Ground truth: Annotations are optional but required if detection is enabled.

  • Formats: Images should be in standard formats (PNG, JPEG, TIFF, …). Ground truth must be in PASCAL VOC XML, default format used with Yolo and Roboflow.

Important

There must also be as many RGB images as SWIR images. For ground truth, there are two possibilities: either the number corresponds to the number of RGB-SWIR pairs, or there is a single ground truth, in which case the code considers the scene to be fixed and uses a single ground truth for the entire execution of the code.

Warning

If files are not paired correctly, the fusion pipeline will fail or produce meaningless results. Always verify that filenames match across visible/ and swir/.

Tips

Tip

  • Keep datasets small when testing new parameters (sql mode).

  • Use consistent naming conventions (e.g. img_###.png).

  • Store large datasets on HPC storage and mount them into the container using -B /path/to/data:/VISWIR/data.