# 5. Fill Pick a backend, press **Fill frames between labels**, and every frame from your first observation to your last gets a position. Your labelled frames always come back exactly as you placed them — a fill never overwrites a label or a detection, and never feeds on the previous fill. Frames outside that span are left empty. (target-labelled-span)= ## Everything happens between your labels The span between your outermost **observations** — labels, plus anything a detector found — is the boundary for the whole workflow, not just for the fill: - **Filling** runs in the gaps *between* labelled frames and nowhere else. This is what makes partial labelling well-defined. If only part of a recording interests you — one bout, one trial, the minute the animal is actually in frame — label its start and its end and work inside it; the rest of the video stays `NaN` rather than being padded with extrapolated positions you would then have to detect and discard. The trade is the obvious one: you get results exactly where you put labels, so a stretch you never bracketed produces nothing at all, however long you spend filling. ## Choosing a fill backend | Backend | How it works | Needs | Cost | |---|---|---|---| | **Spline** (default) | Monotone cubic (PCHIP) interpolation per keypoint over its own labelled frames[^pchip] — geometry only, the pixels are never read | nothing extra | instant | | **Optical flow** | Pyramidal Lucas–Kanade, tracked forwards and backwards across each gap[^lk] | nothing extra | ~video speed, CPU | | **PosePAL (CoTracker3 + refinement)** | A transformer point tracker[^cotracker] whose per-keypoint appearance features are first fitted to *your* labels on *this* video {cite:p}`pan2025posepal`, then tracked forwards and backwards across each gap | torch + cotracker, **GPU** | a few minutes to fit, reused by every fill made from the same labels | Only PosePAL needs installing; see {ref}`PosePAL install `. **Without a GPU**, start with the spline — it costs nothing and is hard to beat when motion is smooth and your labels are dense — and switch to optical flow when the path between two labels isn't a smooth curve, like a fast turn or a wingbeat. **With a GPU**, use PosePAL {cite:p}`pan2025posepal`. It handles longer gaps and larger displacements than either of the others, and because it learns what *your* keypoints look like in *this* recording, it stays on the right leg and the right animal. ## Fit and track with PosePAL PosePAL fills in two phases, and it is worth knowing which one you are paying for: - **Fit** — optimise CoTracker3's per-keypoint appearance features against the frames you labelled. Minutes on a GPU. It depends on nothing but your labels and the video. - **Track** — run the fitted tracker forwards and backwards across each gap. This is what produces the filled frames, and it is fast. **Fill** always does both, but it skips the fit while the fit it already has was made from exactly the labels you have now — held in memory and cached beside the video as `