How it works
Take a photo (or upload one you already have), then press "Read the board". The tool first evens out the lighting — dividing out shadows and glare so the ink stands out from the board — then runs OCR over the cleaned image and records where on the photo each word was found.
Those words are then grouped back together: words sitting close to each other become one block, words across the board become a different block. That's the step that turns three hundred loose words into six notes. Each block appears next to the photo with its outline drawn over the original, so you can check any block against what was actually written and fix it in place.
When the blocks look right, open the result as a canvas. Every block becomes a note positioned the way it was on the board, in roughly the size it was written and, where the photo makes it clear, the colour of marker it was written in.
How to photograph a whiteboard so the text is readable
The single biggest factor is glare. Whiteboards are glossy, so a ceiling light or a window reflecting off the surface produces a white patch that destroys everything under it — no amount of processing recovers writing the camera never captured. Move a step to the side so the reflection slides off the area you care about, or turn off the light directly above the board. If the room has a window along one wall, shoot from the window side rather than towards it.
Stand square to the board and fill the frame with it. A photo taken from a sharp angle leaves the far side of the board smaller, fainter and slanted, and OCR reads slanted text much worse than straight text. Get close enough that the writing is a decent size in the frame — a wide shot of a whole meeting room with a board in the background will not read, because the letters are only a few pixels tall.
Keep the camera steady and let it focus before you shoot. Motion blur and soft focus both smear the thin strokes of a dry-erase marker into grey mush. Use the full-resolution photo rather than a screenshot of the photo or a version compressed by a chat app — resending an image through a messenger typically throws away exactly the fine detail the engine needs.
If the board is very wide, take two overlapping photos and run them separately rather than one distant shot of the whole thing. Two readable halves beat one unreadable whole.
Marker colours and what survives the trip
Black and dark blue markers photograph best: high contrast against a white board, and thick enough strokes to survive compression. Red and green usually read fine too, though red often comes out with lower contrast than it looks to your eye. Light colours — yellow, orange, light green, pale pink — are the ones to avoid. They can be perfectly legible in the room and nearly invisible to a camera, and they are the most common cause of a section of the board simply not appearing in the result.
Where a block's writing is clearly one colour, that colour is carried through to the canvas, so a note written in red stays red and the visual grouping you used in the room isn't lost. When the photo isn't clear enough to call, the block comes back in black rather than guessing.
Drawings, arrows, boxes and connectors are not recognized — OCR reads text, not diagrams. What comes across is the writing, in the position it was written. A boxed-and-arrowed flow will arrive as the labels in the right relative places, and redrawing the arrows on the canvas takes a few seconds once the words are already positioned.
What OCR can and can't read on a whiteboard
The engine here is Tesseract, and Tesseract was trained on printed text. That's the honest framing for everything it does with a whiteboard. Neat block capitals — the way most people write headings on a board — come back well. Careful lowercase printing comes back mostly right, with a handful of characters to fix. Hurried handwriting comes back patchy. Cursive largely does not work at all, and no setting changes that; joined-up letterforms are a genuinely different recognition problem from printed characters.
Numbers and short codes are the least reliable part of any OCR result, on a whiteboard more than anywhere. Ones and sevens, zeros and Os, fives and Ss all trade places, and unlike words there's no dictionary to catch the error. Read every figure before you rely on it.
The tool tells you when it's unsure: any block the engine had low confidence in is flagged "check this" so your attention goes to the parts most likely to be wrong. Every block is an editable text box sitting beside the photo for exactly this reason. Correcting a few words is the expected path, not a sign something went wrong.
If the handwriting doesn't read at all
Sometimes the answer is that the writing on this particular board isn't recognizable, and it's better to say so than to hand back nonsense. When that happens you still have two useful options. The first is to retake the photo with the lighting and framing fixed — glare and distance cause more failures than handwriting does, and a second photo from a better spot often works when the first didn't.
The second is to use the tool as a positioning aid rather than a transcriber: add blocks by hand and type what the board says while looking at it. You still get a canvas with the board's structure on it in a couple of minutes, and it's still faster than laying the whole thing out from scratch. The grouping control also helps here — if the recognized words came out chopped into too many pieces or lumped into too few, regrouping is instant and doesn't re-read the photo.
Why the positions matter more than the text
A plain OCR tool gives you the words in reading order, top to bottom. For a document that's exactly right. For a whiteboard it destroys the information: the whole reason people use a board instead of a notepad is that space carries meaning. Related things are near each other, the timeline runs left to right, the parked ideas are off in the corner, the decision is boxed in the middle. Flatten that to a list and you've kept the words and thrown away the thinking.
That's why this tool returns a canvas rather than a text file, and why the output genuinely doesn't make sense as a .txt download. The blocks land where they were, at roughly the scale they were written, so the board is still readable as a board — and now it's a board you can keep editing, add to next week, and search six months later.
Why nothing is uploaded
Most "whiteboard scanner" tools send your photo to a server, run recognition there, and send text back. That's a normal architecture, and it also means a third party briefly holds a picture of whatever your team wrote on the wall — which for a lot of boards is the most sensitive thing anyone will write down that week.
Here the OCR engine downloads into your browser once and does the reading on your own machine. The photo is decoded, cleaned, recognized and discarded locally; no network request ever carries the image anywhere. The only thing that leaves your browser is the finished canvas, and only at the moment you choose to open it in OmniCanvas.