Near-eye Display Optic Deficiencies and Ways to Overcome Them

Achieving high picture quality, optical fidelity, and natural visual experience is a challenge in near-eye display design.

Learn:

  • Why the conventional approaches do not work for near-eye display optical design;
  • What specific methods exist, and why they are still fundamentally limited;
  • How the computational optical correction techniques allow overcoming those limits.

Read a comprehensive article by Almalence CTO Dmitry Shmunk, originally presented at SPIE AR/VR/MR 2021 conference: http://almalence.com/doc/SPIE-11765-23-V3/Near-eye_display_optic_deficiencies_and_ways_to_overcome_them_SPIE-11765-23-v3.pdf

An ideal imitation of the human eye enables the precise measurement of near-eye and head-mounted display quality

The all-new 2021 version of the Almalence Human Eye Simulator. Optically clear. Eye Tracking ready.

To assess the quality of head-mounted displays, it is necessary to capture images which exactly match with what a human eye would perceive. Indeed, a capturing device has to be capable of accurately replicating the human eye’s optical properties. If they are not, then this could lead to some drastic irregularities – a mismatch of entrance pupil diameter, for instance, would lead to quite different blurs and aberrations, or even sometimes visible Fresnel rings, which are not apparent to the human eye. Once you do have a capturing device in place that can match the optical properties of the human eye, however, then now comes the real challenge: the device has to be recognized as an “eye” by eye trackers – otherwise, there will simply be no chance of capturing a correct picture, as a wrong picture would be displayed in the first place, in case the HMD uses eye position-dependent rendering techniques like foveated rendering or dynamic aberrations correction – which have recently been becoming standard for high-quality near-eye displays.

Almalence, a pioneer in designing the eye-imitating cameras, has now begun to roll out an all-new and updated version of its powerful eye simulator, better than ever and ideally suited for near-eye display picture capturing and quality measurement tasks. It features made-to-order, optically clear eye corneas, flawlessly creating a perfect match to the form of a real, human eye that is indistinguishable for eye trackers – unlike other, off-the-shelf parts and solutions, which commonly result in deviations from the proper shape. A clear aperture for up to a 120° FOV enables the seamless capture of up to the entire field of view in one shot, without ever compromising the contrast and MTF of the true visible picture.

The ideal profile of both visible and IR light absorption and reflection is painstakingly implemented, in order to make the simulator’s iris look exactly like a natural iris to an eye tracker. An additional IR-cut filter also prevents unwanted reflections from the camera lens, which may spoof eye tracker readings.

Auto-focus capability, to avoid having to manually adjust the focal point when moving the “eye” inside head-mounted displays that exhibit a significant field curvature.

The platform encapsulates multiple capturing camera designs, including a 100° field of view camera which enables the user to capture the entire visible FOV in one shot. This feature is also quite useful for geometry distortion measurements. Specially designed narrow 78- and 34-degree FOV cameras are also included, engineered for high-precision optical measurements, including apparent resolution, chromatic aberrations, and more.

A monochromatic camera can also optionally be used with the eye simulator, in order to resolve ambiguity in color channel mixing between the HMD display and CFA filters inside the camera.

Almalence has also developed a powerful software for the processing and transforming of captured images, so that they can readily be used for correct measurements of geometry, MTF, channel crossing, and other quality characteristics, with industry-standard tools such as ImaTest. Together with a 6-DOF robo-arm and its controlling software, all of the above-mentioned features seamlessly combine to present a complete, easy-to-implement tool for head-mounted displays, picture quality assessment, and the profiling of geometry and aberrations correction.

Almalence SuperResolution now supports Mediatek APU

We are happy to announce that our SuperResolution is now fully ported to Mediatek APU, an AI processor powering high-end Mediatek smartphone chipsets.

With the addition of the APU version, Almalence SR now supports the high-end chipset DSPs of both major chipset makers, Qualcomm and Mediatek.

Running SuperResolution image processing on a specialized DSP brings the following advantages:

  • Much faster processing, up to real-time at video frame rates
  • Some 10-20 times lower power consumption
  • CPU offload, no blocking of UI and other applications
  • More sophisticated processing within a shorter time

In 2021, Almalence SuperResolution will be used to achieve unprecedented camera zoom quality on top smartphones powered by the newest Qualcomm 888 and Mediatek Dimensity chipsets.

Microsoft vs Almalence SuperResolution Zoom

While the number one thing that differentiates Microsoft’s new productivity device, the Surface Duo, is the new form factor, we are mostly interested in its camera performance, namely, zoom capability. Having a single camera module, the Surface Duo is said to have enhanced zoom quality by using a super-resolution zoom algorithm.

We were eager to compare Microsoft vs. Almalence Super Resolution Zoom performance, and here it goes – below are comparison results at 7x zoom (max available zoom factor on the Duo).

Siemens Star test chart, 7x zoom, Left: Microsoft Zoom, Right: Almalence SuperResolution Zoom

We particularly like Spilled Coins (aka Dead Leaves) chart, for it is nearly impossible to fake high resolution with it using sharpening. It also extremely well exposes detail loss due to noise reduction, highlighting the advantage of algorithms that suppress the noise without loss of detail.

DxOMark chart, Spilled Coins, 7x zoom. Left: Microsoft Duo built-in, Right: Almalence SuperResolution

Text readability test usually works well too:

Left: Microsoft Surface Duo 7x zoom; Right: Almalence Super Resolution Zoom

To summarize: while using a single camera module seems to make sense for the Surface Duo niche, Microsoft could have used a more powerful super-resolution technique to mitigate the device’s zoom quality limitation.

Huawei P40 Pro Super Resolution Zoom: How Good It Is and How Much Better It Could Be

The Huawei P40 Pro looks like the winner in the smartphone zoom race, achieving DxOMark Zoom Score of 115, the all-time high. However, its outstanding camera hardware, utilizing a 125mm f/3.4 telephoto module, is not the only key to success. Achieving the best results would not be possible without using a computational super resolution zoom technology. We did some testing to check how good it is and if better results could be achieved if using the leading Super Resolution technology from Almalence.

We will start with a side-by-side comparison and then discuss some interesting features of Huawei’s SR which we found during the testing. For the testing, we captured:

  • several JPEG images with the built-in camera app at 10x, those came out pretty different so we used the best one for comparison;
  • a series of RAW images with the 5x telephoto camera module, which were then processed with 2x Almalence Super Resolution.

The pictures were captured indoor, in good office lighting (~700 Lux). Note, as we used RAW images for processing, the colors in Almalence SR output are somewhat off.

Comparing the ability to resolve fine details shows a dramatic improvement when using Almalence Super Resolution Zoom:

Left: Huawei P40 Pro built-in 10x zoom; Right: Almalence Super Resolution Zoom

A strange effect in the next example, most of the fine text is “washed out” in the P40 Pro image. That can be caused by extreme noise filtering or input frames misalignment/deghosting.
Also note the highlighted character. It looks like Huawei’s algorithm employs a kind of a neural network, which tried to “guess” the object but in this case made a wrong guess. (We will show more examples of that NN’s job below)

Left: Huawei P40 Pro built-in 10x zoom; Right: Almalence Super Resolution Zoom

A testing with a wedge chart, Almalence SR Zoom increases the effective resolution by ~20..25% more than Huawei’s built-in algorithm:

Left: Huawei P40 Pro built-in 10x zoom; Right: Almalence Super Resolution Zoom

Getting back to the P40 Pro’s [supposedly] neural network, an interesting example below. First of all, the NN did an absolutely fantastic job resolving the hair (look at the areas 1 and 2). This looks like something beyond the normal capabilities of super resolution algorithms, which makes us convinced a neural network was involved. Exploring the image further, however, we can see that in some areas (e.g. area 3) the picture looks very detailed but actually unnatural (and yes, different from the original), so the NN made a visually nice, but actually a wrong guess. In the area 4, the algorithm “resolved” the eye in a way that it distorted the eyelid and iris geometry, making the two eyes looking at different directions; it also guessed the bottom eyelashes in a way that they look like growing from the eyeball, not the eyelid, which looks rather unnatural.

Left: Huawei P40 Pro built-in 10x zoom; Right: Almalence Super Resolution Zoom.
Huawei’s result looks more detailed, however in some areas those details are unnatural and do not reflect the original object.

To summarize, while the Huawei P40 Pro is clearly the winner in telephoto camera module hardware design, its computational zoom algorithm is not yet doing the best possible job. While having some advantages over Almalence’s Super Resolution Zoom in resolving certain kinds of objects, it could be better in terms of overall resolution capability. It would be really interesting to see what those algorithms could do if combined together, likely that would make an all-time best digital zoom technology.

A Zoom Technology Missing from iPhone 11 Pro

Despite having a telephoto camera module, iPhone 11 Pro zoom is still far behind the top performers which use Super Resolution Zoom.

Zoom has recently become one of the most important features of smartphone cameras with the leading OEMs advertising their devices achieving high picture quality at sometimes crazy zoom levels.

As every high-end smartphone, iPhone 11 Pro uses a dedicated telephoto camera module to achieve the maximum zoom quality. It appears however, that simply utilizing a telephoto module, even of a great design and quality which is undoubtedly the case with an Apple’s product, is not enough to achieve the top zoom performance. According to the DxOMark benchmark, iPhone 11 Pro achieves Zoom Score of 74 while, for example, Xiaomi Mi 10 Pro hits 110, a drastic 1.5x difference!

To go beyond the camera hardware capabilities, top Zoom performers utilize a computational imaging technique, Super Resolution Zoom. As its name suggests, it uses super resolution technique to increase the resolution of the images suffering from the lack of pixels in case the target zoom level exceeds the optical zoom of the telephoto module.

For example: zooming 4x with a 12 MP 2x telephoto module uses only 1/4 of its sensor, or just 3 Megapixels.

Besides improving the resolution, Super Resolution Zoom also increases the SNR, lost due to small aperture of a telephoto module, the higher the optical zoom level – the smaller is the aperture.

We made a few tests to check how Almalence Super Resolution Zoom, the most advanced digital zoom technology, would improve iPhone 11 Pro zooming capabilities. Check a couple of examples below:

iPhone 11 Pro, 4x zoom. Left: iPhone as is, Right: with Almalence Super Resolution Zoom

iPhone 11 Pro, 4x zoom. Left: iPhone as is, Right: with Almalence Super Resolution Zoom

The pictures speak for themselves. Apple can definitely achieve better zoom picture clarity by utilizing a computational super resolution technology.

Almalence Digital Lens to harness the full potential of Varjo’s human-eye resolution head-mounted display

From its beginning Varjo positioned itself as the leader of VR head-mounted displays megapixel race. Their “human-eye resolution” VR-1 truly shows more detail than any other existing HMD.

However, just offering a high pixel count does not mean the user will be able to see a crisp and clean picture through the HMD optics, and in fact, it is the optics which present a display quality bottleneck. In a head-mounted display there are severe design constraints especially in making the optics light weight and fit in a tight space. Those constraints lead to a compromised optical performance, resulting in color fringing and blur. Moreover, any movement of the eye pupil, which itself constitutes a lens element, makes the entire optical system quite different from the original optical design, so the blur and color fringing get even stronger as the eye looks off the optical axis.

Almalence Digital Lens is a computational lens aberrations correction solution which overcomes these limitations by compensating the aberrations of an HMD optics. It does the job of a corrective lens element which dynamically adjusts its properties depending on the eye pupil position. We were eager to check how that technology can improve picture clarity of the highest resolution HMD.

Note 1: This testing was performed by Almalence independently from Varjo. The Digital Lens test was implemented as a Unity application using public API.

Note 2: This is the very first testing, definitely showing sub-optimal results. We see a clear way to further improve the image clarity with the given headset.

We used a construction drawing as a test picture as it clearly demonstrates how the insufficient apparent resolution and clarity limit the VR usability.

The test picture

To take the images within the HMD, we used our camera system with our eye imitator, allowing to capture what a human eye would see.

In the first example the eye looks about 10 degrees off the center. The left part of the gaze area falls onto the high resolution “focus display”, the right part falls onto the lower resolution “context display”

Move the slider left/right to see the difference. Left: Varjo VR-1 as is; Right: VR-1 with Digital Lens. Despite the high display pixel count, the picture does not look very clear. One can even start feeling sick when trying to read the numbers. The very same display with the Digital Lens delivers much clearer and readable picture.

In the next example the eye looks straight at the center, along the optical axis – the ideal case in which the HMD delivers its highest possible picture quality. The gaze area is completely over the focus display. Even in that case the Digital Lens shows a noticeable improvement:

Left: Varjo VR-1 as is; Right: VR-1 with the Digital Lens. Same display, but more legible text and crisper lines.

The beauty of the Digital Lens solution is that it is a pure computational technique, adding no extra size, weight or mechanical complexity to the device.

As mentioned above, those are very first tests, more to follow. However the tests already prove that the Digital Lens is an indispensable technology for high-end VR headsets, allowing to harness the full potential of high display pixel count and density.

Google Super Resolution Zoom: Good Start but not There Yet

Our first testing of Google’s super resolution zoom recently announced in Pixel 3 shows that it indeed can restore some image details, but is still behind the best in class solutions.

Comparing to the “normal” digital zoom, which is basically an upscaling plus edge enhancement, Google’s zoom reveals some details that are indistinguishable in the “normal” image:

However, it’s still not the best of what super resolution can achieve. Below is a comparison of Google’s super resolution zoom to Almalence SuperSensor, a technique based on multi-frame super resolution, running on the same Pixel 3 smartphone:

It looks like Almalence SuperSensor is closer to be delivering on the “optical zoom” promise.

Another nice example captured when taking zoomed images of a book on optical design. Google super resolution makes the text somewhat better readable, however some characters still remain distorted beyond any possibility to recognize them. Also some minor color artefacts are introduced on the originally black text:

And again, if you really want to capture a readable text, Almalence SuperSensor is a solution (note that SuperSensor also got the white color of paper right):

Unfortunately, there is no way to reliably measure the Google’s super resolution processing time. It looks like the processing is performed in background. When trying to quickly open the image right after its icon appears in the camera app, the preview would still show a progress sign for a fraction of a second, probably implying that the image is still being processed. One can feel that the processing takes roughly one second altogether, but there is no way to verify that number. Almalence SuperSensor processing takes 200-500 ms on the same hardware (SnapDragon 845).

The full images used to make the above comparison examples are available in an archive below. Note: for accurate comparison the images of each scene were taken with the same Pixel 3 smartphone, from the same position under the same lighting. The zoom level might slightly differ between images as there is no way to precisely set the zoom factor in the Pixel 3 camera app. The Pixel 3 camera app, updated to the recent version as of Jan 25, 2018, was set to “HDR+” mode.

PS Stay tuned, we will soon show a comparison of super resolution zoom at video frame rates!

Google Super Resolution Zoom: Good Start but Long Way to Go

Please read the updated post.

After publishing this original post we were contacted by Google engineers who pointed out that HDR+ Super Res was not always on in the images we took. Of course, it happened unintentionally, there’s just no obvious way to make sure that Super Res kicks in when taking an image.

We have re-taken the test images as suggested by Google and we admit the results are much better now. However Almalence SuperSensor is still superior :).