In single-camera NIR systems with a filter that blocks all red or blue, the camera can be adjusted to emphasize the color channel capturing NIR. This is done with a custom white balance setting and allows the photos to be directly converted to a facsimile of NDVI. If the custom white balance is not done, the values in the NIR channel will not be high enough to compute directly to meaningful NDVI. This white balance adjustment is a clever hack, and is generally not referred to as calibration.
To calibrate the NDVI image derived from one of these photos, some additional information is needed about the sensitivity of the camera to visible versus NIR light. Typically a target of known reflectance in the bands used to compute NDVI is included in each photo (or representative photos) and used to adjust the values in each pixel so those values will be more closely related to the amount of energy (radiance) in the specified color bands reflected from the scene. That adjustment is the calibration.
Calibration can be done on jpeg photos, but some inscrutable adjustments are made by the camera when if produces jpegs, and these can be difficult to account for. So camera raw photos generally allow more accurate calibration.
Any single camera NDVI system will have NIR and visible light mixed in one of the channels used to compute NDVI. This is a separate problem and can be accounted for by making some assumptions about the amount of contamination. To avoid this complication, a two-camera system can be used so pure NIR and pure visible photos are captured. The two photos can be independently calibrated, aligned (rectified) with each other, and then the NIR and visible values in each pixel can be used to compute NDVI.