The Vision Engine Misreads My Slate Cards
Our vision engine reads slate cards with high accuracy on well prepared cards. If you are seeing many corrections in the review queue, work through this list to find out why.
1. Read the slate conventions page
Most accuracy issues come down to slate cards that are hard for any human to read either. Start with Slate Card Conventions. Compare your worst recent slates to the examples there.
2. Check the image resolution that gets sent
By default the agent downscales every image to 1600 pixels on the long edge before sending it for recognition. This is plenty for a slate that fills a quarter of the frame, but tight for a slate that fills only a small corner.
You can raise this in config.yaml:
vision:
resize_long_edge: 2400
A higher value costs more credits per read. 1600 is the sweet
spot for most rigs. 2400 is right when slates are small in
frame. Going above 3000 rarely helps.
3. Confirm you are on the right model
In config.yaml:
vision:
model: standard
If you have switched to the smaller tier (standard is the
default; high is available for the toughest slates), accuracy
will drop a bit on handwritten cards but cost less per read. Switch back
to high for production runs where every misread is expensive
to chase down.
4. Use the context hint
The vision request includes a small "previous scene and take" context. The
model uses this to break ties when handwriting is ambiguous (the difference
between 012 and O12, for example).
The context comes from the previously processed image. If your team sometimes shoots out of order, you can disable context to remove the bias:
vision:
use_context: false
5. Look for specific common misreads
| Pattern | Likely cause | Fix |
|---|---|---|
Scene O12 instead of 012 | Slate has a cursive zero, no slash | Cross your zeros, or pre-print scene field |
Take 1 instead of 7 | Continental seven without crossbar | Add a crossbar to the seven |
Lens 35mn instead of 35mm | Lowercase double m blurs into n | Print "MM" in upper case |
| Notes mostly missing | Notes field is small relative to slate | Make the notes line at least as tall as the scene and take values |
6. Use corrections aggressively
Every correction in the review queue is folded back into the read history. After 20 to 30 corrections from the same camera assistant's handwriting, the agent starts compensating automatically.
7. Re-run a single image
If you have one stubborn read you want to retry without changing settings, open it in the review queue and click Re-run vision. The model uses the latest correction history, so a re-run after a few corrections often gets it right.
8. Raise the review threshold
If you would rather review more entries than miss any, lower
review.min_confidence in config.yaml:
review:
min_confidence: 90
Anything below 90 confidence then lands in the queue for your eyes.