Converting markdown to ICML ignores all captions. I'm not familiar enough with ICML to know whether this is due to a lack in the ICML spec, but losing information on the way doesn't seem like the best way to implement the conversion.

If the ICML spec allows it, the perfect solution would be to retain the captions.

An acceptable solution would be to convert the text into a paragraph after the image, so it could be converted into a caption manually. Now all converted files have to be manually compared to their source to figure out whether the captions have been dropped or whether the images didn't have captions to begin with.

Right, most other writers print out the caption of figures (i.e. images that form their own paragraph). Implemented it like this as well.