Hi-Lo Card Counting Trainer

Capture

Detector view

Table overview context

Watch the whole table. Nothing here is counted.

no source

Share a screen and it feeds both panels — then drag here to box the whole table.

Drop zone detected here

The area where dealt cards come to rest.

no source

Share a screen, then drag here to box just where the cards land.

no region yet

Share once and both panels use it — then drag out a different region on each: the whole table on the left, just the landing area on the right. Detection only ever runs on the drop zone region, and the tighter you box it the faster and more accurate the read. If the two are on different monitors, share the second one separately from its own panel and each keeps its own screen.

Learning

Teach mode 0 learned

Taught so far

Kept between runs.

Every card that lands appears here with the picture captured of it, and the percentage is how sure the read is. Set the dropdown to what the card really is — that corrects the count and teaches the detector. Correcting a wrong read teaches it more than labelling a skipped one. Amber maybe X? tiles were seen clearly but not named confidently: they are not counted until you say what they are.

no hands yet A page per hand — closed off when the table is swept.

Nothing dealt yet. Share the drop zone and cards will show up here as they land.

Running count 0

True count 0.0

Decks remaining 8.0

Cards seen 0

Dealt 0.0 of 8 decks shuffle in 4.0

Shoe

Decks Shuffle after Rounding

Unseen cards Tray override

Reset automatically at the cut card

Manual entry — keyboard: A 2-9 0 J Q K

Dealt 0

Detector tuning

Frames / sec 10 How often a frame is analysed. Higher costs CPU. Ink threshold 0.16 Raise it if the felt lights up; lower it if faint indices vanish. Min glyph height 8 Ignore ink smaller than this (px in the analysed crop). Rank indices on a live stream are often only 10-14px. Set it to the size of the index — suit pips are admitted down to 45% of it automatically, since they only need to be located, not read. Max glyph height 40 Ignore ink bigger than this. Bracket the rank index tightly. Depth / perspective 1.00 How big a rank index at the top of the region is compared with one at the bottom. The table runs away from the camera, so the dealer's cards are smaller than the players' — 1.00 means no allowance and their indices fall under the size floor. Auto-size measures this for you. Dealer band 30% How much of the region, from the top, holds the dealer's cards. The relaxed size floor applies across this band and tightens below it, so the dealer's smaller indices get through without letting small debris in at the player end. Only has an effect when Depth is below 1.00. Card tilt allowance 45° How far off square a card may lie and still have its suit pip recognised as its own. The dealer's card often lies at a steeper angle than the players'. Too generous and a neighbouring card's pip gets picked up instead, giving a nonsense tilt. Widest character 2.4x How wide a rank may be compared with its height. Upright characters are taller than wide, but a round one — a Q, or the 0 of a 10 — on a tilted, foreshortened card comes out squat and wider than tall. Too low and those vanish before anything can read them. Card brightness 140 How white the area around a glyph must be to count as a card face. Card saturation 90 Upper limit on colour around a glyph. Keeps felt and chips out. Frames held still 5 How long a card must sit in one place before it counts. This is what stops cards being read mid-deal. Position tolerance 0.40 How far a glyph may sit from where it was, as a fraction of its height, and still count as the same card. Low = cards in motion are ignored. Raise it if shaky footage keeps resetting cards. Rank votes 3 Confident reads needed to settle on a rank. Match distance 0.26 How far a glyph may be from a template. Higher = more reads, more errors. Match margin 0.035 How clearly the best rank must beat the runner-up. Re-count lockout 2.5s After a card leaves, how long that area stays locked against counting the same rank again. Analysis width 960px Width the region is scaled to before analysis. Set it above your drop zone's own width and small cards are magnified — that adds no detail, but it lifts the dealer's tiny indices over the size floor and steadies how they rasterise. Too low and the indices shrink below detection, the usual cause of finding nothing. The size window is measured in these pixels, so re-run Auto-size after changing it. Drop the pip under each rank (superseded) The old geometric pip filter. Leave it off: it discards the pip before anything can read the card's tilt from it, and that tilt is what straightens the glyph and tells a top-left index from an upside-down one. Pips are removed by shape instead, after they have done their job. Ignore solid shapes (suit pips) Rank characters are made of strokes; suit pips are filled silhouettes. Measured on real frames, ranks fill 0-6 of 24 rows and pips 9-12. Works whatever angle the card lies at, unlike the geometric pip rule. Straighten slanted characters Cards rarely lie square on, and a slanted A collapses into much the same shape as a slanted 4 — diagonal, crossbar, two descenders. The rank-to-pip vector is the card's own "down" direction, so each character is rotated upright by it before being compared. That also means a rank you teach at one angle still matches at another. Only count cards whose orientation is certain Cards overlap, so one of the two indices is often hidden. A single index still counts — but only when its own suit pip is visible, because that is what tells a 6 from an upside-down 9. Without a pip there is no way to know which way up a lone index is, so it is sent to the learning strip instead of guessed at. Stricter: demand both of a card's indices A stray mark or felt text cannot fake a card if it has to appear twice in the right geometric relationship, and the two readings must agree. But any card lying partly under the next one shows only one index and is missed entirely — usually too costly on a real table. Use the pip to find the top-left index The bottom-right index is the same rank printed upside down, and an inverted 9 is indistinguishable from a 6 on its own. The suit pip settles it: below the rank on the top-left index, above it on the bottom-right. Works even when the card's other corner is hidden. Merge a card's two corner indices A card prints its rank in both opposite corners, the second upside down. When cards lie spread out both are visible, which counts one card twice — and an upside-down 9 reads as a 6, a different Hi-Lo tag. Keeps the top-left one. Also match upside-down glyphs For tables where some cards face away from the camera.

Cards are read where they land. Nothing is counted while a card is in motion — a track only forms once a glyph holds one position, so a card being dealt is ignored until it comes to rest.

Rank only. A card is identified purely by its number or letter. Suit pips are discarded geometrically (the pip sits under the index) and by shape. Hitting ✕ on an amber tile teaches that shape as a non-rank.

It learns as you go. Everything you label is kept between runs, so the detector gets better at your table each session. If it reads nothing at all, open the detector view — that tells you whether it's failing to find glyphs or failing to recognise them.