I’ve employed a work-around I’ve used in the past a few times, and I was able to process your dataset (mostly) successfully!
The images were pretty washed out, low sharpness, and had very low scene detail.
So, I pre-processed them lightly in XNConvert with an Enhance Detail and Auto-Contrast filter.
Original Image:
Pre-Processed Image:
The difference is subtle, but it is enough to really help during the matching phase.
Processing using pretty much default settings worked great:
Options: crop: 0, debug: true, dsm: true, verbose: true
Orthophoto:
DSM:
Point Cloud:
The Point Cloud is where we see some of the limitations of this dataset, this pre-processing workflow, and my settings come together.
The strong “waffling” is just a consistent lack of tiepoints extracted during matching. Maybe we could do significantly better by forcing the min-num-features up, or by cranking up pc-quality, feature-quality, and turning off resizing.
To that end, Piero is tracking a feature request for OpenDroneMap to do this pre-processing in the background transparently for the user just for the matching phase, and then pass the original/unprocessed images to the pipeline for point-cloud coloration, orthophoto generation, and model texturing.