Per-Box Settings Reference

Each box has its own settings. Select its field in the field list (or its box on the picture), and the Box tab of the inspector, on the right of the window, shows them under the field’s name, in three sections:

Reading: what the box contains (Type, Format) and how the value is written.
Filtering: which readings are accepted.
Image cleanup: how the box’s picture is prepared for a Tesseract model. It is collapsed until you open it (ScoreSight remembers whether it’s open), and its header says how many of its settings you changed (all default, 2 changed…) and has a Binary view button that shows the result. While a PP-OCR model reads the boxes, its settings are disabled (and kept) with the note Image cleanup is for the Tesseract models; PP-OCR reads the picture as it is: PP-OCR reads each box’s picture untouched, which works much better for it. They stay available for a box that counts lights.

Reset to defaults puts every setting of the field back to its default. Every slider shows its value next to it (Off when the setting is off), and every setting has a tooltip, which also gives its name in ScoreSight 0.1.x. Settings are saved with your layout and included in File > Export field layout…. In exported files and scoresight.json, they appear under each box’s settings object, using the JSON keys listed here.

A few settings apply to every box at once. They are in All boxes, at the bottom of the Box tab, always available: the OCR model, Detections / s (see Performance) and Update on change.

For a task-oriented guide, see Recognition Tuning.

On this page

Summary
Processing order
Reading
Filtering
Image cleanup
Template
1. Build from other fields
Not shown in the window
1. OBS source name
2. is_custom
Other controls

Summary

UI label (0.1.x label)	JSON key	Values	Default
Reading
Type	`type`	`0` Number, `1` Time, `2` Text	Time box: Time. Others: Number
Format (Select Preset)	`format_regex`	regular expression	depends on the field (see below)
Remove leading zeros (Remove leading 0s)	`remove_leading_zeros`	on/off	off
Add ordinal suffix (Ordinal)	`ordinal_indicator`	on/off	Period: on. Others: off
Count lights (Dot Counter)	`dot_detector`	on/off	off
Filtering
Minimum confidence (Conf. Th)	`conf_thresh`	0.00–0.99 (slider 0–99 = the OCR model’s confidence in %)	0 (off)
Smooth over last 5 readings (Average Output)	`smoothing`	on/off	off
Skip OCR while image is unchanged (Skip Similar Image)	`skip_similar_image`	on/off	off
Image cleanup (Tesseract models and the dot counter)
Black & white (Binarize)	`binarization_method`	`0` Whole picture, `1` None, `2` This box, `3` Adaptive	`0` Whole picture
Remove specks (Cleanup)	`cleanup_thresh`	0.00–0.99 (slider 0–99)	0 (off)
Squash height (V.Scale)	`vscale`	1–10	10 (no scaling)
Straighten slant (Skew)	`skew`	−10 to 10	0
Thicken strokes (Dilate)	`dilate`	0–5	1
Invert colours (Invert Input)	`invert_patch`	on/off	off
Crop to text (Auto Crop)	`autocrop`	on/off	off
Resize to 35 px (Rescale Input)	`rescale_patch`	on/off	on
Normalize digit width (Normalize W-H Ratio)	`normalize_wh_ratio`	on/off	off
Per-character sub-boxes (Composite)	`composite_box`	on/off	off
Skip empty values (Skip Empty Values)	`skip_empty`	on/off	on
Template
Build from other fields (Template Field)	`templatefield`, `templatefield_text`	on/off, text	off, empty
(not shown)	`is_custom`	true/false	true for fields you added with +
OBS page: Text sources	`obs_source_name`	text	see OBS source name

Processing order

With a Tesseract model, for each box, on every detection cycle, ScoreSight:

takes the part of the box that is inside the processed picture (a box that reaches past the edge is read from its visible part; its saved position isn’t changed),
turns it into black and white (Black & white),
removes small specks (Remove specks),
squashes it vertically (Squash height) and shears it (Straighten slant),
thickens strokes (Thicken strokes),
inverts it (Invert colours),
reports the box as Empty without running OCR if the image is blank (every pixel the same, for example a box over plain background); a Count lights box reports 0 instead,
skips OCR if the image hasn’t changed (Skip OCR while image is unchanged),
crops to the text (Crop to text),
scales it (Resize to 35 px, Normalize digit width),
counts lights (Count lights) or runs OCR with the character set of the Type,
checks the raw text against the Format and the Minimum confidence,
applies Smooth over last 5 readings (to accepted readings only), Remove leading zeros and Add ordinal suffix,
marks the reading Empty if no text is left, or SameNoChange if it equals the box’s last good reading and Update on change is on.

Each box is processed on its own copy of the picture, so overlapping boxes never affect each other. Binary View shows the result of steps 1–5 for every box: exactly the image the OCR receives, before Invert colours.

With a PP-OCR model, ScoreSight skips the Image cleanup steps (2–6, 9 and 10) for every box but a Count lights box, which is processed as above. For the others it:

takes the part of the box that is inside the processed picture, in colour,
reports the box as Empty without running OCR if the picture has almost no contrast (a display that is off, a black frame, a box over plain background: less than 20 gray levels between its darkest and brightest parts, ignoring the 2% extremes), since a model would guess,
skips OCR if the picture hasn’t changed (Skip OCR while image is unchanged; the picture is compared thresholded on its own, so a change is seen on a dim display too),
reads every box of the frame together with the character set of each box’s Type,
continues with steps 12–14 above: Format, Minimum confidence (PP-OCR’s confidence), smoothing, leading zeros, ordinal suffix.

Binary View shows a PP-OCR box’s picture in gray: what PP-OCR reads, without its colour.

Reading

Type

type: which characters the OCR may output.

Value	UI	Allowed characters
`0`	Number 0-9	`0123456789`
`1`	Time 0-9 , . :	`0123456789` plus `:` and `.`
`2`	Text	letters, digits, space and common punctuation

Use: always set the narrowest type. Number for scores, fouls, periods and shot clocks. Time for clocks. Text only for names or other words.

In ScoreSight 0.1.x the Type of a box was lost after a restart or export (#22). It is saved since 0.2.0.

Format

format_regex: a regular expression the raw OCR text must match completely, before averaging, leading-zero removal and ordinal suffixes. If it doesn’t match, the reading is marked FailedFilter (yellow) and outputs that send only good values ignore it.

The Format list shows the preset whose expression is in the field under it, or Custom for any other expression. Choose a preset to use its expression, or type your own (choosing Custom selects the expression for typing). An expression that isn’t a valid regular expression gets a red border and the message Not a valid expression, so it is ignored: the OCR then accepts every reading. Its tooltip says what is wrong, for example missing ), unterminated subpattern.

Preset	Expression	Default for
Time mm:ss.d	`mm:ss` or `ss.d` (expression below)	Time
Time mm:ss	`^[0-5]?\d:[0-5]\d$`
Time ss.d	`^[0-5]?\d\.\d$`
Time 0-59	`^[0-5]?\d$`
Shotclock 0-39	`^[0-3]?\d$`	Shot Clock
Score 1dd	`^1?\d{1,2}$`	Home Score, Away Score
Score ddd	`^\d{1,3}$`
Period 1-4	`^[1-4]{1}$`	Period
Period d	`^\d{1}$`
Alphanumeric	`^[A-Za-z0-9]*$`
Any text	`^.*$`
Any number	`^\d*$`	Home Fouls, Away Fouls, custom fields

The Time mm:ss.d expression is:

^(?:(?:[0-5]?\d:[0-5]\d)|(?:[0-5]?\d\.\d))$

Use: always. A tight format is the most effective protection against wrong values reaching your graphics. Examples of custom formats: ^[0-3]?\d$ (shot clock with or without a leading zero), ^(1[0-2]|[1-9]):[0-5]\d$ (12-minute quarters), ^[0-4]$ (downs).

A brace that doesn’t form a repeat count, as in ^\d{1,2$, is read as a plain character, so that expression is valid but no score matches it: every reading is rejected (yellow). Check the expression when a box only shows rejected readings.

Remove leading zeros

remove_leading_zeros: removes zeros at the start, so 07 becomes 7 (and 00 becomes 0). An empty reading stays empty. In ScoreSight 0.1.x: Remove leading 0s.

Use: for scoreboards that zero-pad scores or fouls, when your graphics should show 7.

Add ordinal suffix

ordinal_indicator: appends an ordinal suffix: 1 becomes 1st, 2 becomes 2nd, 3 becomes 3rd, 4 becomes 4th. In ScoreSight 0.1.x: Ordinal (1st, 2nd, ..).

Use: for periods, quarters, innings or downs when the graphics should show “2nd”. On by default for Period.

Count lights (dot counter)

dot_detector: instead of reading characters, counts the separate bright blobs (larger than about 5 square pixels) in the processed box image and reports the count, for example 3. The reading is always considered valid, and a box with no lit blobs reports 0. In ScoreSight 0.1.x: Dot Counter.

Use: for indicator lights such as balls/strikes/outs, possession arrows, bonus lights or timeouts left. Use one box per row of lights, and check in Binary View that each lit light is a separate white blob. Use Invert colours if lights appear dark.

Filtering

Minimum confidence

conf_thresh: the minimum confidence, in percent, that the OCR model must report for a reading. The slider goes from 0 to 99, and the setting is stored as a fraction (slider ÷ 100, so 60 % is stored as 0.6). A reading below the threshold is marked FailedFilter (yellow). 0 turns the check off, and is the default. In ScoreSight 0.1.x: Conf. Th.

With a PP-OCR model, the confidence is the mean probability of the characters read (0–100 %). It is reliable: in the OCR evaluation, with a threshold of 90 % PP-OCR (accurate) still published 85 % of its readings, and only 1 % of the boxes got a wrong value, against 4 % with no threshold.
With a Tesseract model, ScoreSight uses Tesseract’s mean confidence for the recognized text. Tesseract sometimes reports 0 for a correct, non-empty reading; in that case the average of the per-character confidences is used instead.
A value above 1 in a hand-edited settings file is taken as a percentage (60 means 60 %).

Use: with PP-OCR, 80–90 rejects most wrong readings; the box keeps its last good value meanwhile. With Tesseract, leave it off unless wrong values get through a strict Format, then try 30–50 and raise it slowly while watching for correct readings that turn yellow: Tesseract’s correct readings of real scoreboard footage often score only 30–50 %, so high thresholds reject good values. With either engine, a box whose readings never reach the threshold never updates, so set it while watching your scoreboard.

In ScoreSight 0.1.x this setting had no effect, because a confidence of 0–100 was compared with a threshold of 0–1. ScoreSight 0.2.0 compares them correctly. Since the old default (50) never did anything, boxes saved by an older version get their threshold set to 0 (off) once, the first time 0.2.0 loads them, so they read exactly as before. The settings file records this with "boxes_schema_version": 2, and the log mentions it.

Smooth over last 5 readings

smoothing: replaces each reading with the most common value among the box’s last 5 accepted (Success) readings, comparing whole values. Ties go to the most recent value. Rejected and empty readings don’t enter the history. In ScoreSight 0.1.x: Average Output.

Use: for values that flicker between two readings, for example during LED refresh. A new value must win the vote before it is shown, which delays changes by a couple of detection cycles, so prefer a stricter Format or better exposure for fast clocks.

ScoreSight 0.1.x voted character by character, so a change from 12 to 9 could briefly produce 92. 0.2.0 votes on whole values.

Skip OCR while image is unchanged

skip_similar_image: compares the processed box image (with PP-OCR: the box’s picture, thresholded on its own) with the image of the box’s last accepted (Success) reading. If fewer than 2% of the text pixels (the digits, not the background) changed, OCR is skipped and the box repeats its last good value as an unchanged reading (SameNoChange). The comparison image isn’t refreshed on skipped cycles, so slow changes add up until they are read. It isn’t refreshed by a rejected or empty reading either, so when the first reading of a new value fails, the box is read again on the next cycle. Resizing the box starts over. In ScoreSight 0.1.x: Skip Similar Image.

Use: to save CPU on values that change rarely (scores, periods). Because the change is measured against the digits, a one-digit change (for example 18 to 19) is read. Turn it off if a very small change (a single segment of a large display) is ever missed.

ScoreSight 0.1.x measured the change against the whole box, so a one-digit change could be skipped, and it reported skipped cycles as a yellow SIM reading. It also compared with the image of the last reading, even a rejected or empty one, so a new value whose first reading failed wasn’t read again until the image changed once more.

Image cleanup

These settings prepare the box’s picture for a Tesseract model, and for the dot counter (Count lights) with any model. While a PP-OCR model reads the boxes they are disabled, and kept for when you go back to Tesseract.

Black & white

binarization_method: how the box is turned into black and white before OCR. In ScoreSight 0.1.x: Binarize, with the choices Global, No Binarization, Local and Adaptive.

Value	UI (0.1.x)	How it works	Use when
`0`	Whole picture (Global), the default	One automatic (Otsu) threshold computed over the whole picture	The picture is mostly the scoreboard (crop helps) and lighting is even
`1`	None (No Binarization)	The grayscale image is passed to the OCR as is	Clean, high-contrast screen captures and TV graphics
`2`	This box (Local)	An automatic (Otsu) threshold computed for this box only	Dark surroundings, small board in a wide shot, low contrast. A good first thing to try
`3`	Adaptive	A threshold that varies across the box (Gaussian adaptive)	Uneven lighting, glare, reflections or gradients across the digits

The whole-picture threshold is computed only when at least one box uses Whole picture, so the other methods don’t pay for it.

Remove specks

cleanup_thresh: removes connected blobs smaller than a fraction of the box area. The limit is (slider position ÷ 100) × 5% of the box area, so at the maximum, blobs smaller than about 5% of the box area are removed. In ScoreSight 0.1.x: Cleanup.

Use: for specks, dust, LED “dead pixels” or bits of the frame inside the box. Too high a value erases thin digits such as 1.

Squash height

vscale: squashes the box vertically to value/10 of its height (10 = unchanged; the slider shows 100 % to 10 %), keeping it centred. It has no effect on boxes only a pixel high. In ScoreSight 0.1.x: V.Scale.

Use: for very tall, narrow fonts that the OCR misreads. Try 60–80 %.

Straighten slant

skew: shears the image horizontally by value/40 (−0.25 to +0.25) to straighten slanted digits. In ScoreSight 0.1.x: Skew.

Use: for italic fonts or a camera that is slightly rotated. Adjust until the digits stand upright in Binary View.

Thicken strokes

dilate: thickens the bright strokes by this many passes of a 3×3 dilation (0 = off). It applies to the image the OCR reads with every Black & white choice, including None. In ScoreSight 0.1.x: Dilate.

Use: raise it when the segments of 7-segment or dot-matrix digits don’t touch each other. Lower it to 0 when digits merge into blobs.

In ScoreSight 0.1.x, Dilate changed the Binary View but didn’t reach the OCR with every binarization method. In 0.2.0 what you see in Binary View is what the OCR reads.

Invert colours

invert_patch: swaps black and white before OCR. In ScoreSight 0.1.x: Invert Input.

Use: when readings improve with the colours swapped, or for dark digits on a light background combined with Count lights.

Crop to text

autocrop: crops the box to the area that contains bright pixels, plus a 10-pixel margin, before OCR. With display style All, the cropped area is shown as a green rectangle. In ScoreSight 0.1.x: Auto Crop.

Use: when the box must be larger than the digits, for example because the value changes width. Combine with Remove specks so specks don’t widen the crop.

Resize to 35 px (recommended)

rescale_patch: scales the box image to 35 pixels high before OCR, the size the bundled models work best with. In ScoreSight 0.1.x: Rescale Input.

Use: keep it on. Turn it off only to experiment with external models trained at another size.

Normalize digit width

normalize_wh_ratio: measures the typical width-to-height ratio of the digits (the median over the last 50 accepted characters 0 and 2–9), then stretches or squeezes the image horizontally so digits come out about half as wide as they are tall. The measurement starts over when you change Black & white, Squash height, Straighten slant, Thicken strokes or Crop to text, and it isn’t saved: it is measured again after a restart. In ScoreSight 0.1.x: Normalize W-H Ratio.

Use: for unusually narrow or wide fonts.

Experimental and legacy settings

The last two settings of Image cleanup, under Experimental / legacy, don’t change how the box is read.

Per-character sub-boxes

composite_box (with the per-character sub-boxes stored as mini_rects): experimental. Lets you draw one sub-box per character. The sub-boxes are saved, but in this release recognition still reads the whole box (not used for reading yet). See Composite boxes. In ScoreSight 0.1.x: Composite (Per-Character).

Skip empty values

skip_empty: never write an empty value to the box’s text file or OBS text source. In ScoreSight 0.1.x: Skip Empty Values.

Use: none any more (no effect since 0.2.0), it’s kept for one more release. In ScoreSight 0.2.0, text files and OBS text sources receive only good readings, and an empty reading is never a good one, so they keep showing the last value when the board blanks out for a moment (a flickering LED, a person walking in front of the board, a camera flash), whether this option is on or off.

Template

Build from other fields

templatefield, templatefield_text: makes a custom field a template whose value is built from other fields, for example {{Home Score}} - {{Away Score}}. Insert field next to the template adds a field at the cursor, spelled exactly as the template needs it. Available for custom fields that aren’t on the picture (a standard field shows only custom fields can be built from other fields). See Template fields. In ScoreSight 0.1.x: Template Field.

Not shown in the window

OBS source name

obs_source_name: the OBS text source this box updates (see OBS Studio). Set it in the Text sources table of the OBS page (Outputs tab > OBS). Standard fields use Home score, Away score, Clock, Period, #Home Fouls, #Away Fouls and shotclock. Custom fields have an empty name, and an empty name doesn’t update OBS. (In 0.1.x it could only be changed in an exported configuration file.)

is_custom

is_custom: true for fields created with +. Custom fields can be renamed, deleted and turned into template fields.

Other controls

Reset to defaults (in ScoreSight 0.1.x: Defaults), at the top of the Box tab, or Reset settings to defaults in the field list’s context menu: restores all settings of the selected box to the defaults for its field.
Remove from picture also resets them: the next time you place the field, its box starts with the defaults. When you changed any of its settings, ScoreSight asks first.

The Force Format checkbox of ScoreSight 0.1.x had no effect and was removed in 0.2.0.