Processing forest images with ODM remains difficult. I have now switched to using the WebODM Lightning server (command line option) with greater success. I find that simply using the defaults (min-num-features 8000) and using
./odm images/image_folder_name --skip-3dmodel
works well most of the time. I still have some occasional problems with receiving only partial orthoimages from some image sets, and the stitching process is very sensitive to the image overlap. At an altitude of 150 m, I am unable to obtain consistent orthoimages with overlaps of less than 80% for forests, especially after the canopy closes. The best overlap appears to be near 90%.
When I was using ODM with a virtual server, I also found that rerunning the program with the same image set often produced different orthoimages (sometimes complete). It appears that the selection of strips for image matching is random.
Because I have had difficulty in obtaining consistent orthoimages from my flights, I have had to develop R scripts to analyze and align individual images. (I use a Sentea Double 4K Multispectral Sensor to obtain narrow band red, green, and blue bands from one camera and red edge and near infrared from a second camera in the sensor.) Looking at overlapping images, I am amazed that ODM works at all. The parallax effect on adjacent images is visually striking and images of closed canopies are difficulty to align manually. There is indeed magic in the ODM algorithms.
Regarding the pmvs option. I have found that the documentation of options lags code changes in the various versions of ODM and keeping current with updates is thus a high priority. I now exclusively use the Lightning network WebODM for my work, and I recommend using the command line to insure consistent use of options. I have found that using drop down menus in WebODM interface to manually select options for each job is an open invitation for error.