Quick Answer: ScanDot is an experimental Android 3D scanning app. It uses ARCore camera tracking and depth estimates to capture point clouds on supported phones, saves them locally, and lets you rotate them in an interactive viewer. The output is a sparse, noisy point cloud, not a textured mesh or a measurement. A separate workstation pipeline handles video photogrammetry with FFmpeg and COLMAP.
If you want a 3D scan from your phone without a studio rig or a LiDAR sensor, an Android 3D scanning app can be worth a look. ScanDot is open source on GitHub and still early. It doesn't promise a finished model. It shows you the raw point cloud, which is more useful if you know what to expect before you start.
What ScanDot Actually Captures on Your Phone
ScanDot uses ARCore Depth, Google's depth API, to estimate how far surfaces are from the camera. As you move the phone, the app combines camera tracking with those depth estimates and builds a cloud of points. Each point is a position in space. There's no triangulated surface and no colour texture mapped onto it.
Here's how a session typically goes. Name the scan, tap Start scan, and move slowly around a subject such as a coffee mug on a desk. Work in good, even light. Save the scan, then open it from Results. The scan lives in the app's private storage, so it doesn't appear in your photo gallery.
Device support is the first hurdle. ARCore Depth runs only on a subset of Android phones, and Google publishes a compatibility list at developers.google.com/ar. Check that list before you spend time on a capture, because a phone that runs ARCore for AR apps may still lack the depth support ScanDot needs.
The practical rule: slow movement and steady light matter more than the number of passes you make. Fast sweeps cause tracking drift, and drift shows up as a smeared or doubled cloud.
Why Full-Depth Mode Can Make Surfaces Look Worse
The viewer has an optional full-depth mode. Most people assume more depth data means a more accurate model. The README says otherwise: full-depth mode interpolates missing depth values, and that interpolation can distort surfaces.
Here's the cause-and-effect chain. ARCore returns no depth for reflective or low-texture regions. Full-depth mode fills those holes by estimating values from neighbouring points. On a flat tabletop, that's fine. On a curved object with a gap in one area, the fill can bridge across the edge and pull a surface into the empty space, so a rounded handle looks flatter or stretched.
So the default view is the safer choice when you're checking geometry. Turn on full-depth mode when you want a denser visual impression, and treat its shapes as a guess. The trade-off is simple: fewer invented points means more holes, and more points means more invented ones.
This next part matters more than it looks. The viewer shows whatever the phone measured, and it doesn't tell you which points were measured and which were filled. Compare the default and full-depth views of the same scan before you trust either one.
The Video Pipeline Runs on Your Workstation, Not Your Phone
Many people assume the app turns ordinary video into a 3D model on the phone. It doesn't. The repository includes a separate Python pipeline that runs on a computer, and the README is explicit: the Android app does not reconstruct ordinary video with photogrammetry on the phone.
The workstation pipeline has three stages:
- Extract frames. FFmpeg and FFprobe pull still frames from your video. The benchmark's
preparecommand takes an--fpsvalue, and the README's example uses--fps 5. - Run sparse reconstruction. COLMAP estimates camera poses and a sparse point cloud from those frames. COLMAP implements the Structure-from-Motion and Multi-View Stereo methods described by Schönberger and Frahm in their 2016 CVPR paper, Structure-from-Motion Revisited.
- Run dense reconstruction (optional). Adding
--denseasks COLMAP for a denser cloud. It needs a compatible COLMAP configuration, and the README warns that a successful run doesn't prove the geometry is complete or accurate.
The commands look like this:
python -m processing.benchmark prepare "path/to/input.mp4" --output "path/to/new-run" --fps 5
python -m processing.benchmark reconstruct "path/to/new-run" --colmap "path/to/colmap" --dense
Two details trip people up. The output directory must not already exist, so the command fails if you reuse a run folder. And you need Python 3.12 or newer. Omit --dense for a faster, sparse-only run, then decide whether the dense pass is worth the extra time. Run --help on each subcommand for the full option list.
Comparing Phone Capture With Workstation Reconstruction
The two halves of the project solve different problems, and choosing the wrong one wastes time. This table sums up the differences from the README.
| Approach | Where it runs | Input | Output and limits |
|---|---|---|---|
| In-app ARCore capture | Android phone | Live camera plus ARCore Depth | Saved point cloud; gaps and drift on reflective or textureless subjects |
| In-app full-depth view | Android phone (viewer) | Saved point cloud | Interpolated depth; can distort curved surfaces |
| COLMAP sparse run | Workstation | Frames from FFmpeg at --fps 5 | Sparse camera poses and points; no dense surface |
| COLMAP dense run | Workstation with compatible COLMAP setup | Same frames, --dense flag | Denser cloud; completeness and accuracy not guaranteed |
Use the phone when you need a quick capture of a small object and can walk around it. Use the workstation when you already have video and can spare the compute time. If you need measurements, neither path is validated yet, so don't use them for engineering decisions.
Where Scans Fail: Reflective, Transparent and Textureless Subjects
Most scan failures aren't random. They follow from how depth and tracking work. Here are the failure modes that matter most.
- Shiny metal and glossy plastic. Specular reflections confuse depth estimates, so the cloud has holes or false surfaces behind the reflection.
- Glass and clear containers. The depth sensor sees through the glass or to its edges, which leaves ghost points.
- Plain white walls and matte grey objects. Without texture, tracking has little to latch onto, and drift creeps in.
- Long walks around a large subject. Small errors accumulate over the loop, so the start and end of the sweep don't line up.
Here's the counter-intuitive part. Scanning a richer, more textured object doesn't always give a better result than a plain one. A patterned mug may track well and still produce a noisy cloud if the reflections are strong. The useful test is whether the subject has stable, matte, visibly textured surfaces, and whether the light is even.
Most people stop here and blame the app. Don't. Check your lighting, your speed and your subject before you decide the tool is broken.
Is This Android 3D Scanning App Worth Trying Yet?
Yes, if you treat it as a prototype. The project has three honest limits to keep in mind. The README says results can contain gaps, drift and noise, and that they are not textured meshes or validated measurements. Capture quality and device compatibility still need broader testing. And the CI pipeline builds the app and runs unit tests, but it doesn't exercise a camera, ARCore tracking, GPU reconstruction or measurement accuracy.
You have two ways to try it. The first is the preview APK, whose release notes describe compatibility and signing limitations. The second is building it yourself. You need JDK 17, Android SDK 36 and an ARCore Depth-capable device. From the android-app folder, run ./gradlew assembleDebug lintDebug testDebugUnitTest (use gradlew.bat on Windows).
Install the debug APK as an update rather than a fresh install, so your existing scans survive. Uninstalling removes them permanently, so export anything you care about first. The project wiki covers installation, capture and troubleshooting in more detail.
Frequently Asked Questions
What is ScanDot and how does it relate to an android 3d scanning app?
ScanDot is an experimental open-source Android app for capturing 3D point clouds on supported phones. It uses ARCore Depth for depth estimates, saves scans locally and offers an interactive viewer. A separate workstation pipeline handles video photogrammetry with FFmpeg and COLMAP.
How do I 3d scan with an android phone using ScanDot?
Check that your phone supports ARCore Depth, then name the scan and tap Start scan. Move slowly around the subject in good, even light, save it, and open it from Results. Avoid fast sweeps, because they cause tracking drift.
Why does my 3d scan have gaps or look distorted?
Reflective, transparent and textureless surfaces return little or unreliable depth, which leaves holes. Full-depth mode fills those holes by interpolation, which can distort curved surfaces. Try the default view, improve your lighting, and rescan the problem area more slowly.
Can this android 3d scanning app turn a regular video into a model on my phone?
No. The app doesn't run photogrammetry on video. To reconstruct a video, you run the Python pipeline on a workstation: FFmpeg extracts frames, and COLMAP performs sparse or dense reconstruction. The output directory must not already exist.
Are scans from this app accurate enough for measurements?
Not yet. The project describes its results as early prototype output with possible gaps, drift and noise, and not validated measurements. Use scans for visual inspection or experimentation, and don't rely on them for dimensions until you've tested them against a known reference.
Pick a small, matte object on a plain surface, scan it once in default view and once in full-depth view, and note where the two disagree. That single comparison will teach you more about the Android 3D scanning app than any spec sheet. Pass the preview APK to a friend who's struggling with reflective objects and ask them to try a matte test subject first.