Slate Card Conventions
Slate Tracker reads slate cards using our vision engine. The model is very tolerant, but a few simple habits will push the accuracy from "very good" to "we never need to correct anything."
What works best
Contrast
Use a dark marker on a white or off white card. Black marker on white card stock reads best. Avoid red or blue on dark cards; the model gets there but confidence scores tend to be lower.
Clean printed labels
If you can pre print scene, take, lens, and ND fields, do it. Hand fill the values. Printed templates beat hand drawn templates by a measurable margin in our tests.
Block lettering
Hand printed block letters beat cursive every time. If your team has one person whose handwriting reads more reliably, ask them to fill the cards when you can.
Steady, square framing
A slate held flat, parallel to the lens, fills more of the frame and reads more reliably than a slate held at an angle. The camera does not need to be close, but the slate should not be skewed past about 20 degrees.
Single slate per frame
If multiple slates are visible (an A camera and B camera slate in the same frame), the model has to guess which is canonical. Try to keep the secondary slate covered, or shoot a single slate frame at the start of each setup.
What to include on the card
Slate Tracker reads any printed or written text on the card. The fields the agent looks for, in order:
- Scene
- Take
- Lens
- Notes (free text; goes into the dashboard entry)
If a field is missing, the agent fills it with unknown and
flags the entry for review.
What hurts accuracy
- Reflective laminated cards under direct sunlight. The glare can obscure whole fields.
- Slates held in motion. Even small motion blur on the text drops confidence sharply.
- Cards with multiple corrections layered on top of each other. The model can see the layers and may read the wrong one.
- Slate text smaller than 50 pixels tall in the final image. If you shoot 4K and crop, this is rarely an issue. If you downscale before processing, keep the slate readable.
A working example
Here is a slate card our test set reads at 99 percent confidence:
+----------------------------------+
| SCENE : 012 |
| TAKE : 4 |
| LENS : 35mm |
| NOTES : interior, day, kitchen |
+----------------------------------+
The same card with cursive scene and take numbers reads at around 87 percent. The same card under harsh window glare reads at around 70 percent.
When to correct vs reshoot
If confidence is above 80 and the dashboard preview looks right, accept it. If confidence is below 80, click into the entry, verify the values, and correct anything wrong. Corrections improve future reads on slates with the same handwriting.