Learning goal
After watching, you can diagnose the six most common MaskAnyone failure modes — install errors, container crashes, memory limits, detection failures, tracking loss, segmentation glitches — and apply the right fix without needing developer help.
Jump to a symptom
Transcript
Cold open
Something broke. Container won't start, mask jumps off the body, GPU isn't being used. Six fixes, five minutes. Find your symptom, jump to the timecode, you're back in business.
1. "Cannot connect to Docker daemon"
If you see this error, Docker isn't running.
On Mac and Windows: launch Docker Desktop. Wait 30 seconds for it to come up — the whale icon in the menu bar should stop animating. On Linux:
sudo systemctl start docker
If it still won't connect on Linux, check systemctl status docker to see what the service is complaining about.
2. "Port already in use"
If you see "address already in use" on port 8080 or 443, something else is using that port.
Two fixes. Either stop whatever's already on that port — lsof -i :8080 will tell you what — or edit docker-compose.yml and map MaskAnyone to a different port like 8081.
3. Containers keep restarting
If a container keeps cycling — restarting, exiting, restarting again — look at its logs. That's where the real error is.
docker-compose logs maskanyone-backend
Nine times out of ten on Mac or Windows: Docker doesn't have enough RAM. Open Docker Desktop → Settings → Resources → Memory. Bump it to at least 8 GB. 16 GB if you can.
4. "Out of memory" during processing
Different from the install-time OOM. This one hits mid-job, usually on a long high-resolution video.
Three fixes. One: split the video into shorter segments and process them separately. Two: re-encode to a lower resolution if your analysis tolerates it. Three: move that specific video to a beefier machine — cloud GPU by the hour is roughly the cost of a coffee.
5. Detection / tracking failures
Detection failure: the model didn't notice someone in the frame. Common with small distant figures, dim lighting, weird camera angles. Add them by hand. Click on the person in their first visible frame — the manual prompt becomes the seed for tracking the rest of the video.
Tracking loss / identity swap: the model loses someone behind an obstacle and reassigns the wrong colour when they reappear. Re-identify them in the frame they reappear. Click; the colour locks back on. Then re-process only the affected segment.
6. Sloppy mask boundaries
Segmentation glitch: the mask boundary is wrong — chops off a hand, leaks past the body, misses a piece of clothing. Click on the boundary and drag. Then re-process. If a specific type of glitch keeps recurring — same kind of clothing, same camera angle — that's a sign your source video quality is the bottleneck, not the model.
7. GPU isn't being used
If your processing speed is way below what your hardware can do, you probably installed MaskAnyone but the GPU isn't wired into Docker.
Test it:
docker run --rm --gpus all nvidia/cuda:11.0-base nvidia-smi
If this command fails, your container can't see the GPU. Go back to the installation deep dive, section on the NVIDIA Container Toolkit.
Recap
Six symptoms. Docker daemon — start Docker. Port collision — stop the conflict or remap. Container restart loop — read the logs, probably memory. Mid-job OOM — segment, downsize, or move to bigger hardware. Detection or tracking failure — click to fix in the editor. GPU missing — verify with nvidia-smi inside Docker.
If your problem isn't one of these six, open an issue on the MaskAnyone GitHub repo. Paste the relevant chunk of docker-compose logs — that's the fastest way to get help.