Ongoing ClearType Experiments

From VISTA LAB WIKI

Jump to: navigation, search

Contents

[edit] Open Questions (High Level)

Question: What, if any, is the relationship between delta E signal strength and delta E area? Is there a function that captures people's perceptual judgements between those two properities? I.E. is a TAS delta E for a peak chromatic error (say 2d gaussian with sigma 0.2) perceptually equivalent to a TAS delta E where the peak is delta E low but spread over a large area (say 2d gaussian with a sigma of 5).

Question: What is inter-letter variability in delta E score within a display?

Question: What is inter-letter variability in delta E scores between displays?

Question: What is inter-letter variability due to initial subpixel starting position, e.g. start drawing letter aligned on a red sub-pixel versus starting the drawing aligned on a green sub-pixel?

Question: What is inter-display variability in 1D filters? I.e. should different filters be used on different displays?

Question: What effect does viewing distance have on 1D filters? Within and across displays

Question: Evaluate 2D filters, are they better or worse than 1D filters? Lots of test permutations, e.g. filter parameters, font size, sub-pixel alignment, display type, viewing distance, character, dpi, etc.

Question: What is distribution of deltaE score over a single letter? I.e. are specific visual features in letters causing high/low delta E scores? Could possibility find with removal evaluation, i.e. systematically remove sections of a letter and check how that impacts upon delta E scores.

Question: Are there minimum tweaks to letters to have a significant impact upon delta E scores? E.g. turn on 5 extra pixels leads to large drop in delta E with minimum change in letter shape (compare downsampled letter shape with highres letters shape, e.g. early compare in ScoreFontMultiResolutionChange.m?)?

Question: Can we identify problem features by analysising the deltaE error maps? i.e. find regions which 15% worst scores and see how they correlate with letter features, subpixel layouts and filters. Are bubbles for letter features the same as rendering quality features (I doubt it but worth checking, and even if so then why - just measure of "feature" complexity, i.e. harder to draw features have more visual "information / features")? Will this work where we have already tried to optimize the letter rendering via filters? We may need to try to optimize the rendering because the errors over a whole letter may not be localised to specific features.

Question: (Thoughts after trying variations of TAM deltaE) - mean deltaE over any area throws away information about the spatial size of the perceptual color difference. Should the size of the +1 deltaE area by used to weight the deltaE error measure?

[edit] Open Questions (Low Level)

Question: How do letter shape and size effect deltaE mean calcs? Do cross compare of mean, min, max, percentage, sum and cropped / no-cropped of calculating deltaE over letter. Think letter shape and size interaction with mean calc (effectively smoothing based on letter size / shape??).

Question: Are there hand annotated image distortion maps for fonts we're testing? Like in "Color Image Fidelity Metrics Evaluated Using Image Distortion Maps". Could be useful when thinking about features which matter for rendering letters on screen.

Question: What would image distortion maps by channel look like? Potentially interesting because of: Individual channel analysis.


[edit] Notes

In the questions above, deltaE is interchangeable with blur or other metrics.

Mean delta E is used to calculate letter error in ctFontCompare(). Is that good or bad? A mean calc ignores maximums and minimums in delta E scores and looks at deltaE over the full letter. Are perceptual color artifacts a problem for peak deltaE values rather than means, i.e. specific parts of a letter have peak deltaE which corresponds to a perceptual artifact which human observers would see?


[edit] Archived Results

Results pre-2011 Microsoft status report



[edit] Test (Ongoing): Evaluating SCIELAB, ctToolBox And DeltaE's

Method:

Display simulated:

Fonts simulated:

Font sizes:

Viewing distance:

Script (generate data):

Script (analysis):

Results:


[edit] Test 3 (Phase 1 - Ongoing): How Does Colorblindness Interact With ClearType Filters?

Purpose: Test how does colorblindness interact with cleartype filters, i.e. should different cleartype filters be used if a person is colorblind?

Method: Simulate various types of colorblindness, then test what a+b 5 tap fitlers are generated and how deltaE score changes.

Display simulated:

Fonts simulated:

Font sizes:

Viewing distance: 0.60 (ammended - version 2)

Status: Need to fix whitepoint calcs


[edit] Test 2 (Phase 4): Rendering On Calibrated Displays

Purpose: Get calibrated displays working, and generating Threshold Maps. As we need this to visually confirm delta E predictions.

Method: Code changes to remove difference map delta E artifacts from calibrated displays

Display simulated: RGB Dell Chevron and Dell Stripes

Fonts simulated: Arial, letter a to z & A to Z

Font sizes: 11pt

Viewing distance: 0.60 (ammended - version 2)

Results:

Threshold Map Dell Chevron 11pt 96 dpi letter lower case w.
Threshold Map Dell Chevron 11pt 96 dpi letter lower case v.
Threshold Map Dell Chevron 11pt 96 dpi letter lower case s.
Threshold Map Dell Stripes 11pt 96 dpi letter lower case w.
Threshold Map Dell Stripes 11pt 96 dpi letter lower case v.
Threshold Map Dell Stripes 11pt 96 dpi letter lower case s.


[edit] Test 3 (Phase 1 - Colorblindness): Evaluating Effects Of Individual Differences On ClearType Filters

Purpose: Check whether differences in color perception, due to colorblindness, has a significant effect on what filters are appropriate, i.e. should differet ClearType filters be used dependencing on whether the viewer is has color deficient vision?

Method: Implemented colorblindness simulation in font rendering pipeline. These results shouldn't be taken as definitive, as this mostly about testing the implementation and calcs.

Display simulated: Theortical RGB, 72 dpi

Fonts simulated: Arial, letter N, 100 dpi

Font sizes: 14 pt

Viewing distance: 0.60 (ammended)

Results: Compared to a normal viewer - at this very very initial stage it seems that the deuteranope has a much lower TAS delta E, tritanopes higher and protanope see the letter with no color artifacts (don't believe that).

Results of finding 1d 5 tap filter for normal color perception.
Results of finding 1d 5 tap filter for deuteranopia color perception.
Results of finding 1d 5 tap filter for tritanopia color perception.
Results of finding 1d 5 tap filter for protanopia color perception.


[edit] Test 2 (Phase 3): Tested effects of drawing simulated 1 subpixel on

Purpose: Are the Threshold Maps generated for simulated displays reasonable when turning off a single subpixel? Very controlled stimuli.

Results: Threshold Maps and corresponding delta Es do seem reasonable, though what to confirm the by-reference stimuli is drawn as black.

Method: Draw 1 subpixel on simulated RGB display. Then generate Threshold Maps with TAS delta E scores.

Display simulated: RGB vertical

Stimuli: Vertical Single 1 subpixel turned off

Viewing distance: 0.60

Threshold: 1 & 0

Script (generate data): Customized version of GenerateSpecialStimuli (using stimuli 2) and seRenderImge.m (tweaked to draw special stimuli on-screen rather than letters)

Results:

Threshold Map for simulated RGB display where 1 red subpixel is turned off. Threshold 1.
Threshold Map for simulated RGB display where 1 green subpixel is turned off. Threshold 1.
Threshold Map for simulated RGB display where 1 blue subpixel is turned off. Threshold 1.
Threshold Map for simulated RGB display where 1 red subpixel is turned off. Threshold 0.
Threshold Map for simulated RGB display where 1 green subpixel is turned off. Threshold 0.
Threshold Map for simulated RGB display where 1 blue subpixel is turned off. Threshold 0.
File:OnScreen-1subpixelONLY-redOFF-RGB-100dpi.png
Image shown on screen when 1 red subpixel is turned off. Note: To see the image properly you need to click on it. IMPORTANT: This image doesn't shown the effect of turning off the single subpixel, as the image compression algorithms get rid of the color artifact. You need to run the matlab script to see the artifact (a tiny light blue dot) - ShowFilteredLetterNoPsych with special stimuli 2.
File:OnScreen-1subpixelONLY-greenOFF-RGB-100dpi.png
Image shown on screen when 1 green subpixel is turned off. Note: To see the image properly you need to click on it.
File:OnScreen-1subpixelONLY-blueOFF-RGB-100dpi.png
Image shown on screen when 1 blue subpixel is turned off. Note: To see the image properly you need to click on it.

[edit] Test 2 (Phase 2): Tested effects of drawing simulated 1 subpixel wide column on TAS delta E score

Purpose: Are the Threshold Maps generated for simulated displays reasonable when turning off a single column of subpixels? Very controlled stimuli.

Results: Threshold Maps and corresponding delta Es do seem reasonable, though what to confirm the by-reference stimuli is drawn as black.

Method: Draw 1 subpixel wide column on simulated RGB display. Then generate Threshold Maps with TAS delta E scores.

Display simulated: RGB vertical

Stimuli: Vertical column 1 subpixel wide.

Viewing distance: 0.60

Threshold: 1

Script (generate data): Customized version of GenerateSpecialStimuli (using stimuli 1) and seRenderImge.m (tweaked to draw special stimuli on-screen rather than letters)

Results:

Threshold Map for simulated RGB display where 1 column of red subpixels are turned off.
Threshold Map for simulated RGB display where 1 column of green subpixels are turned off.
Threshold Map for simulated RGB display where 1 column of blue subpixels are turned off.


[edit] Test 2 (Phase 1): Evaluating Subpixel Positioning

Purpose: Confirm we're drawing characters and other visual stimuli correctly on screen at the subpixel positions we expect.

Results: Subpixels are drawn correctly in scripts. Below are the images generated by turning off the different subpixels.

Method: Generated bitmap for drawing on-screen, where turned off 1 subpixel in turn, e.g. red subpixel, green and blue. Then visually confirmed that the subpixel we expected was turned on and off, this confirmation was done by viewing the screen through a microscope and viewing the individual on-screen subpixel activations.

Display simulated: None, used real display

Stimuli presented: Single subpixel column of subpixels turned on

Viewing distance: Varied (0.60)

Script (generate data): Tweaked version of ShowFilteredLetterNoPsych.m & GenerateSpecialStimuli.m (with stimuli 1

File:OnScreen-1subpixel-vertical-redOFF-RGB-100dpi.png
Image shown on screen when red column (1 subpixel wide) of subpixels are turned off. Note: To see the image properly you need to click on it.
File:OnScreen-1subpixel-vertical-greenOFF-RGB-100dpi.png
Image shown on screen when green column (1 subpixel wide) of subpixels are turned off.
File:OnScreen-1subpixel-vertical-blueOFF-RGB-100dpi.png
Image shown on screen when blue column (1 subpixel wide) of subpixels are turned off.
Personal tools