Review workbench
The review workbench is where a reviewer works through a project one recording at a time. It shows the media, every annotation layer the platform can produce for that recording, and an honest account of what each layer did or did not do here.
Three columns:
- The queue: filters, counts, and one row per recording with its verdict and the layers it carries.
- The stage: the media with every enabled overlay drawn on it, plus the language contract, the label workbench and the readiness receipt.
- The layer rail: one row per applicable layer, with its state, its measurements and its run control.
The layer rail
Every layer that could apply to this recording gets a row, including the ones that did not run. A layer you cannot see is a layer you cannot reason about.
| What the row shows | When |
|---|---|
| The reason this data cannot carry the layer | The layer is inapplicable for this kind of data |
| Turned off for this project, with a link into pipeline settings | The layer’s mode is off |
| Produced for this recording, with its measurements | The layer ran here |
| Nothing produced for this recording yet | The layer is available but has not run here |
| Live progress and a cancel control | A job for this layer is in flight |
Expanding a row separates this recording from across the project, because they answer different questions. For this recording you get the measurements the layer recorded, its receipt notes, the model that produced it and where the output was written. For the project you get how many recordings carry it and the quality signal the layer measured across them.
A layer that measured nothing says so. “This layer produced output here but recorded no measurement” and “not assessable, nothing produced here yet” are different statements, and neither is ever rendered as a zero or a passing score.
Running a layer on one recording
Heavy layers run through the annotation job queue, never as a blocking request, so a slow lane reports progress instead of appearing to fail. The rail shows the queue position, the item in flight and a live count, and cancel is available from queued and running alike. Cancelling stops the lane without undoing it: recordings already finished keep their output, and the job record names what landed.
Re-running is always a decision
A layer that already produced output here never offers a plain run. It offers “run again”, which states that it replaces the existing output and asks you to confirm before anything is queued.
That mirrors the queue’s own rule. Re-queueing a layer over recordings that already carry its output is refused unless you explicitly ask for it, so work you already paid for is never silently redone and never silently skipped. See the already-produced rule.
Overlays
Every layer’s overlay toggles on and off independently, so you can compare two lanes on the same frame instead of switching screens.
| How the layer reaches the stage | Layers |
|---|---|
| Drawn over the media as boxes, tracks or skeletons | Object detection, object tracks, hand pose, body pose |
| A rendered image layer with an opacity control | Semantic segmentation, depth, optical flow, 3D |
| A band under the timeline | Scene captions, action segments, structured metadata |
| A panel below the stage | Dense description |
When there is nothing to draw, the toggle is disabled and says why rather than disappearing: the layer has not run here, or it produced stored output that has no rendered stream yet, or it reads in the panel rather than on the frame.
Semantic segmentation offers per-class visibility when the stored mask carries one class index per pixel. When the stored stream is already colourised, the legend says per-class hiding is not possible on it rather than showing toggles that do nothing.
Overlay colours come from a colourblind-safe palette and follow the light and dark themes.
The language contract
Captions, action segments and dense description are presented as one contract rather than three unrelated panels, because they are three resolutions of the same claim about what happened: one sentence, the steps, and the full account. Each carries its own receipt, and each is editable in place by a reviewer with the role for it.
An edit replaces the machine output rather than sitting next to it. That is why the segmentation metrics score only against a separately stored adjudicated set: scoring the stored output after a review would compare it to itself. See why these need a gold set.
Vocabulary
The workbench also shows what this project’s labels actually say, joined against what the workspace declared: which classes were found, how often, with what properties and values, which declared classes were never seen, and which observed classes were never declared. Renaming or merging a class rewrites this project’s stored labels and tracks and keeps each label’s provenance.
Declaring or undeclaring a class changes the workspace vocabulary that every project in it shares. Renames and merges are local to the project you are in.
Next
- What each layer produces: Annotation layers.
- Turning a layer on or off: Pipeline settings.
- How Euler grades its own annotations: Automated review pass.