Turning Phone Video Into Photos for Photogrammetry and 3D Scanning

Photogrammetry builds a 3D model from lots of overlapping photos of the same thing. Newer techniques such as Gaussian splatting work from the same kind of input. Either way, the software needs dozens or hundreds of sharp images taken from slightly different positions.

You can take those photos one at a time, but filming a slow walk around the subject is easier and makes it hard to leave gaps. You then turn the video into a set of stills, which is frame extraction. This guide covers how to film for it, how many frames to pull out, and the mistakes that make reconstructions fail.

Why video works well for this

The hard part of photogrammetry capture is coverage. Every part of the subject needs to appear in several images, each overlapping heavily with its neighbours. With a camera it is easy to miss an angle and only find out when the model has a hole in it.

With video you move continuously, so coverage is close to automatic. The trade-off is quality per image: a video frame usually has fewer pixels than a photo from the same phone and is more prone to motion blur. Both are manageable if you film with extraction in mind.

How to film it

• Move slowly and steadily. Motion blur is the most common reason frames get rejected, and a steady pace also keeps frames that are evenly spaced in time evenly spaced around the subject.
• Lock exposure and focus. On iPhone, press and hold on the subject in the Camera app until AE/AF LOCK appears. Brightness changing from frame to frame makes matching harder.
• Stop the phone switching lenses. Recent iPhones can jump between lenses as you move closer or further away. Settings, Camera, Record Video, Lock Camera prevents that during a recording.
• Shoot in 4K if you can. More pixels give the software more detail to match.
• Use soft, even light. Overcast daylight is ideal. Hard shadows that move as you move, and shiny or see-through surfaces, are difficult for every reconstruction tool.
• Make loops. For an object, circle it at two or three heights. For a room, walk along the walls facing inward, then across the middle.

How many frames to extract

What matters is overlap. Each image should share most of its content with the next one; 60 to 80 percent is the range most photogrammetry guides recommend. Too few frames and the software cannot connect them. Too many nearly identical ones and processing slows down for no gain, and positions that are almost the same make depth harder to work out, not easier.

A practical way to think about it is angle. Circling an object, one frame for every 5 to 10 degrees gives good overlap, which is 36 to 72 frames per loop. Then work out the count from your video’s length: if one loop took 40 seconds and you want 60 frames from it, that is one frame every two thirds of a second.

Doing it with FrameRipper

1. 1Trim the clip first. FrameRipper spreads frames evenly across the whole video, so the ten seconds of walking up to the subject and fumbling for the stop button would take frames away from the scan. On iPhone, open the video in Photos, tap Edit, drag in the ends of the timeline and save it as a new clip.
2. 2Load the trimmed video and set the frame count from the table above. The line above the controls shows the resulting gap between frames.
3. 3Choose JPEG at quality 92 or higher. Reconstruction software works from detail and texture, and high-quality JPEG keeps both at a fraction of the size of PNG. Two hundred 4K PNGs can come to well over a gigabyte.
4. 4Extract, download the ZIP and unzip it into its own folder.
5. 5Scroll through and delete any frames that are clearly blurred. A handful of bad frames can do more harm than a few missing ones.

Filtering blurry frames automatically

For large scans, deleting blurry frames by eye gets tedious. A simple sharpness score does most of the work: blurred images have little edge detail, which shows up as a low variance of the Laplacian. The dataset guide has a short OpenCV script for exactly this. Run it over the folder and look at anything it flags before deleting.

Feeding the frames to reconstruction software

Most tools take a folder of images, and extracted frames are named in order (frame_0001, frame_0002), which helps. COLMAP, for example, has a sequential matching mode built for video frames: it compares each image mainly with its neighbours in filename order instead of with every other image, which is far faster on large sets.

Meshroom and COLMAP are free and run on your own computer. Several phone apps and commercial tools accept video directly and do the extraction step themselves. Extracting yourself is still useful there: you choose the frames, you can throw out the blurred ones, and you can reuse the same set in different tools.

If the footage came from an iPhone with HDR on, check that the frames do not look flat. Low-contrast images give the software less to match. Why iPhone HDR frames look washed out has the fix.

Scanning rooms and homes

A scan of a room records a lot: the layout of your home, what is on the shelves, sometimes documents or screens in view. FrameRipper never uploads your video, so extracting the frames adds no exposure, but the reconstruction step might if you use a cloud service. Check where your chosen tool processes images before scanning anywhere private.