Merging Projects¶

File ▸ Merge into Project… combines another .slp into the one you have open — for pulling together work split across people, sessions, or machines.

The preview¶

Click Choose .slp files… and pick the file to merge in. Before anything changes, the app matches the two projects and shows a preview: how many Videos and Tracks match or are new, and whether the Skeleton matches. Nothing is applied until you click Merge.

Merging several files¶

You can pick several .slp files at once, for example prediction files from a few inference runs. They're merged one at a time, in the order you picked them, and the dialog shows File 2 of 3 and so on. Each file gets its own preview and conflict review:

• Merge applies that file and moves on to the next.
• Skip this file (or closing the dialog) leaves it out and moves on.
• After the last file, the dialog closes.

Each merged file is its own undo step. A file that can't be read, or whose skeleton differs, can only be skipped; the others still merge.

Videos are matched by basename and tracks by name, so two projects labeling the same videos line up even if the files sit at different paths.

Conflicts¶

A conflict is when both projects have a user-labeled instance on the same frame of the same video, close enough to be the same animal (within 5 pixels). Predictions never count as conflicts. The app groups conflicts into clusters — a base instance can clash with several incoming instances and vice versa, so they're resolved as a unit rather than pairwise.

If there are no conflicts, the dialog says it's a clean merge and Merge combines the two projects.

If there are, pick a Global rule for them:

Reviewing conflicts¶

For anything you don't want decided by the global rule, the conflict list (sortable by Frame, Track, and distance, and filterable) lets you set Keep per conflict: Both, Base, or Donor. Click a row to draw the competing instances on the actual frame so you can see which is right. Reset choices drops your per-row picks and goes back to the global rule.

This is worth doing when two people labeled the same frames — the disagreements are exactly the frames where your labeling guidelines are ambiguous, and seeing them is useful beyond the merge itself.

After merging¶

The result toast summarizes what happened. The merge is a normal command, so Cmd+Z / Ctrl+Z undoes it.

Run Label Quality Check afterwards — merges are a common source of duplicate instances.

Corrections made in a separate predictions file¶

Predictions don't always land in your training project. A predictions .slp made outside the app, or one you opened from a remote job with Open predictions instead of merging it, is a separate project. If you correct predictions there, those corrections are labels, but only in that file.

Training only sees the project it trains on. Training on the predictions file alone would build a model from your corrections without your original labels, so bring them together first:

1. Save the file you corrected.
2. Open your training project.
3. File ▸ Merge into Project… and choose the corrected file.
4. Review any conflicts, click Merge, then save and train.

Only accepted, corrected instances count as labels. Predictions you didn't accept come along in the merge but never train the model.