It’s interesting what can be generated with some random data, or at least data that was not created for the purposes of photogrammetry. The right set of images sourced from multiple places or a sharp enough piece of video can make for some cool experiments in point cloud generation.
To demonstrate this, we grabbed some drone footage posted to YouTube that captured a small section of the White Cliffs of Dover, an 8km coastal stretch of stunning chalk-white cliffs in the south-eastern corner of the United Kingdom.
We downloaded the video, captured by YouTube user VdubPhotogLife shooting from a DJI Phantom 3 advanced, using some simple web-based exporting software searchable on Google.
VLC Media Player was then used to export individual frames from the video. This is done by accessing the advanced settings of VLC and making adjustments to the scene video filter settings, such as how often a frame should be captured, and which format and resolution to capture a frame in.
The scene video filter needs to then be turned on in the filters subheading of video settings.
As the video captured good footage throughout it’s run time we captured two frames every one second for the length of the video and ended up with more than 400 images exported.
Then it was a matter of aligning them in Agisoft Metashape, formerly PhotoScan, and seeing how accurate our results were.
After detecting pairs across the 400+ images Metashape initially recognised about three clusters of cameras, some footage was just too different in colour and exposure to be matched with others. Refining the set of images to a range of about 300, where the image had a relatively consistent exposure provided enough information to generate our sparse and dense point clouds.
Following some cleaning up, a mesh and texture were created. Some additional mesh editing also took place to extend the plane where the ocean would be, as the original capture gave too much point variation over these sections due to moving waves, ocean depth, water reflections etc.
Here’s the final model for inspection: