Review
The review screen is where you look at what Euler made of your data, one episode at a time: one clip with every layer drawn on it, the 3D scene and the depth beside it, the camera’s sensor streams under it on the same playhead, and the evidence behind every readiness check. You accept or reject from here.
Three columns:
- The queue: All, Needs review, Accepted and Rejected, with counts, and one card per episode with its verdict.
- The stage: the player with three tabs (Video, 3D, Depth), the sensor streams under it, then the language panel, the label workbench and the readiness receipt.
- The layers rail: one row per layer; it owns everything drawn on the frame.
On a narrow window the rail starts as a strip of layer icons and opens over the edge of the clip, so the clip is never squeezed; your choice is remembered.
One clean clip, painted per frame
Each episode plays one clean clip. Nothing a model produced is burned into the pixels: hand and body skeletons, object outlines and tracks are painted live on top, and the picture and the skeleton for each frame are painted together, so the skeleton stays on the hands while the clip plays. Poses come from the full-density landmark files; between two close samples the pose is interpolated, and across a real gap nothing is drawn, because a held skeleton would be a guess.
If a clip will not play, the stage says why: the file is missing, this browser cannot decode the clip’s format (try Chrome, Safari or Edge), it could not be decoded, or it could not be loaded, with a retry.
The layers rail
Each row has the layer’s icon, an on/off switch, Solo and, for layers drawn on the frame, a strength slider. A layer with no data on this episode is dimmed. A layer that is turned off for the project links to Pipeline settings; a background layer that has produced nothing here offers Run, which queues it for this episode. A layer that already produced output is never re-run from here; re-running is a Pipeline settings decision, where the cost is shown first.
| How a layer reaches the stage | Layers |
|---|---|
| Painted on the frame | object outlines, object tracks, hand pose, body pose |
| A stored image layer over the clip, at a strength | segmentation, depth, flow, 3D |
| A band under the timeline | captions, sub-tasks, event tags, metadata |
| A panel under the stage | dense actions |
Segmentation can hide classes one by one when the mask stores a class per pixel. Colours follow the light and dark themes and are colour-blind safe.
The 3D tab
A live 3D view of the episode, on the same playhead as the video. Under the 3D layer in the rail, each part has its own switch and strength:
| Part | What you see |
|---|---|
| Scene cloud | the scene from every depth keyframe, edge-shaded so shapes read; in real colour after a 3D rebuild, else coloured by height (or by distance or confidence) |
| Current depth | the depth keyframe nearest the playhead as a surface |
| Camera path | where the camera, the wearer’s head, went; smoothed, and coloured by how confident the tracking was |
| Camera now | the camera at the playhead with the current video frame on it |
| 3D hands | both hands as 21-joint skeletons with a short fingertip trail; marked experimental |
| Object centres | tracked objects placed in 3D |
| Floor grid | one-metre squares, axes and a scale bar |
A part this episode cannot draw stays listed with the reason.
Controls. Drag to orbit, shift-drag to pan, scroll or pinch to zoom. Keys: arrows orbit, plus and minus zoom, C follows the camera, R resets (or double-click), M maximises the 3D view to fill the review column, F goes fullscreen, space plays. Maximised, the view has its own play bar and the video sits in a corner. Save a PNG of the view, or record a 15-second turntable to a WebM file. A head-motion card shows the smoothed path length, head speed and a top-down map of the path. What am I seeing? explains the axes and units.
The Depth tab
The depth keyframe nearest the playhead:
- colour ramps (Turbo, Viridis, Inferno, Euler, grey) with a colourbar in metres and a histogram;
- one range held across the episode, or fitted to each frame;
- hover for the distance at a pixel; click two points to measure between them;
- an RGB wipe and an overlay at an opacity; fullscreen.
If the depth is relative rather than metric, the tab says so and shows no metres. If the estimated focal length disagrees noticeably with the camera’s known lens, a quiet note says so.
Sensor streams
For recordings with camera telemetry, graphs run under the clip with one shared playhead: click or drag to seek; left and right arrows step one frame (shift for one second); plus and minus zoom around the playhead and 0 shows the whole episode. Sub-task and action lanes and the flagged moments sit above the graphs.
| Group | Graphs |
|---|---|
| Motion | accelerometer, gyroscope and its magnitude, gravity tilt, camera yaw, pitch and roll |
| Exposure | shutter, ISO, brightness |
| Derived | head speed and turn rate (estimated), hand presence left and right, hand speed, object count |
Each graph has units and a hover readout, and collapses on its own. A stream the recording does not have shows a quiet “not recorded”. Location and face streams are never shown.
Readiness receipt and QC evidence
For first-person data the receipt opens on a ring that counts the output checks that pass, sit in review, fail, or were not checked, with the readiness score beside it. Then:
- Certificate and routing, side by side: the certificate’s verdict for this episode and the queue’s routing. If they ever disagree, a sentence says why.
- Flagged moments: a strip over the episode marking every stretch a check flagged. Click one to jump the video and the 3D view there.
- Every check, grouped as timing and sync, motion, task and language, and annotation coverage: what it means in a sentence, the measured value against its threshold, a bar showing the pass, review and fail zones and where most of your project sits, the certificate row, and thumbnails of the flagged stretches. A check measured once for the whole clip says so instead of pointing at a moment.
- Not checked: what did not run here, and why.
The capture checks (hands in shot, camera on the head, idle time, lens obstruction, exposure, head motion, dropped frames and the rest) show their own evidence: seekable stretches and the worst frames. What each check means is in Readiness and certificate.
Accept, reject and edit
Accept or reject the episode from the stage; human decisions are never changed by automatic re-routing. Captions, sub-task steps and dense actions are shown as one account at three resolutions (a sentence, the steps, every action) and are editable by a reviewer; an edit replaces the machine output rather than sitting beside it. The label workbench edits object labels, and the vocabulary panel shows which classes the project actually found against what the workspace declared.
Next
- What each check means: Readiness and certificate.
- Export what you accepted: Exports.
- From code: the review routes.