Background

Previously, I've had a lot of difficulty working with the images I was producing from my Sony a6000 and producing accurate SfM models of complex features in vegetation. While the a6000 takes great photos I was having problems reconstructing the fine branches and leaves of targets using Photoscan.

This problem was also evident when I attempted to collect a 100 meter transect on the SRER using the a6000.

My thinking is that the many unmatched photos are caused by the dramatic motion of walking and the changing POV as I moved down the length of the transect or around a target plant or tree. 

I bought a DJI Osmo to test out how stabilization using a 3-axis gimbal, as well as shooting video in 4k improves the collection of data. My hope is that I will have enough images from the 4k video to determine whether the failure to reconstruct branches is simply a parallax problem or if the SIFT algorithm is not making enough matches within the canopy structure of the trees during reconstruction for the 3+ images needed to generate a matching point.

Methods

After setting up the Osmo to run on my Galaxy s7 smartphone (using DJI Go app from Google Play + firm-ware upgrade), I took video of a couple of shrubs and cacti in my yard.

I copied the 4k video files (.mp4) off the micro SD card onto my workstation (I'm using a RAID0 SSD drive to keep the video on).

I used Adobe Creative Cloud Premiere Pro (Media Encoder) to convert the *.mp4 video to .JPEG images at 29.97 frames per second (the native video frame rate) and the highest resolution quality settings.

As a test I tried running the analysis keeping the *.mp4 / *.mov data as well as the exported *.jpegs on the CyVerse Data Store (via Mountain Duck). Unfortuntately it appears that the i/o of Mountain Duck is too slow to keep up with the Media Encoder. 

Steps

Open the Premiere Pro or Media Encoder (Creative Cloud)

Add the .MP4 file to the Projects

Click on File > Export (Ctrl + M)

In Export Settings select:

Format: JPEG

Preset: Custom

Output Name: [Select Folder for Archive]

Results

Because I was using the SSD for multiple file operations (moving large files onto and off the disk) during the video to frame reconstruction the job took over 40 minutes to complete. 

The output images looked pretty good despite the fact they are from a moving video.