Limits and known gaps

What the platform does not do yet, or does with a caveat. We would rather you read it here than find it in your training run.

Data and annotation

• Relative depth. Depth is metric by default, with the camera’s intrinsics estimated from the image. If a project uses a relative depth option, the values have no unit: the Depth tab says “relative depth”, shows no metres, and the export’s depth JSON says so.
• Estimated focal length. On wide-angle and fisheye action cameras the estimated focal length can disagree with the lens; the Depth tab notes it, and distances read from depth carry that error.
• 3D hands are experimental. They are 2D hand keypoints lifted through estimated depth, so they exist near depth keyframes rather than on every frame, and they are not yet at training-grade accuracy. They are marked experimental in review and in the export README, and you can leave them out of any export.
• The camera path is estimated. It comes from the video, the depth and the camera’s own motion sensors. Raw paths include head bob and tracker jitter, so path length and speed are reported smoothed, with the raw value kept beside them.
• Object labels on video carry no frame time yet. Use object tracks for time-placed objects; an object that leaves view and comes back starts a new track.
• The wearer filter can keep a bystander’s hand right next to the wearer’s on a small share of frames.
• Some output checks cannot place their finding in time. They measure one number for the whole clip and say so.

Product and operations

• Background job records do not survive a platform update. Finished work is kept, but a job that was running during an update shows as missing rather than interrupted. Check long jobs afterwards.
• Some re-derivations run while the tab is open (re-measure, 3D rebuild, track clean-up, wearer filter, re-export). The screen sends them a few episodes at a time; closing the tab stops the remaining batches.
• Moving a source works for data uploaded into Euler; bucket sources cannot be moved or shared across projects yet.
• Proxies cannot be forced at ingest from the app.
• Re-route review reports counts, not the list of episodes that moved.
• Large projects: the review queue loads every episode record at once; very large projects are slower to open.
• Browsers: the review player is checked in Chromium-based browsers. A browser that cannot decode H.264 cannot play the clips, and the player says so.