NEW Leica Q3 ISO-Invariance, Dynamic Range, High ISO & OIS

welcome back to my channel for a short Leica Q3 episode and this video is a follow-up to my introductory video about the Leica Q3 which I recently posted actually on the day when Leica announced the new Leica Q3 and you find a lot of details about this camera already in this video so look it up if you have not seen it I'll post the link Down Below in the info box in this video now I want to further explore the new sensor in the Leica Q3 and Leica speaks in the marketing material about a new backside illuminated so acronym BSI 60 megapixel sensor but I'm quite sure the sensor in the Leica Q3 is a very close relative of the sensor in the Leica M11 there are three topics I want to quickly discuss in this video the first is I want to collect more data points on dynamic range and in the context of dynamic range try to find out whether the sensor of the Leica Q3 has a property which people call ISO invariance second I want to shoot the camera with very high eyes or value and see up to which eyes or value the images are still usable when it comes to noise and Clarity of the image and third I want to try out and find and that's a test I typically do with a new camera how long can I shoot handhold before images become blurry so in other words what is the longest exposure time I can afford to do with handhold shooting let's kick off the video [Music] thank you [Music] let's go right away into my three topics so the first one was dynamic range and what I want to do with this camera is I will go into an available light environment namely the main train station in Zurich where ISO values are going up very quickly and will do a series of test shots and then discuss them in Lightroom and come up with some conclusions which might be interesting for you to know and they will also show again that the dynamic range of this new Leica Q3 sensor is just amazing thank you foreign [Music] I want to show now some frames which I shot in the main train station in Zurich and I did shoot them all with the same ISO parameters so this one here for instance was shot with an ISO of 3200 and then I shot all hand hold of course the same frame with an ISO of 200 so four full stops underexposed and if we go back to the first one with the high ISO Valley and I go into the develop section here you see I've not post-proced these images so they are kind of ugly this was really about testing not about providing photography art and in the next frame which then was four full stops underexposed the only slider I used was the exposure slider where I pushed up the exposure by four full stops and then I'm going to compare them now by selecting them both in the library so let's do this here let's get this one let's put them side by side they are not exactly the same frames because they were shot handheld but nevertheless we can compare them and if you look into them these are kind of the same images so even if you look into darker areas there is you know if you have a lot of fantasy a little bit more noise on the one where we did shoot with the low ISO value and pushed up the exposure slider by four full stops but essentially these are the same images they look good and that was my first thought when I saw this this will very likely be an ISO invariant sensor and that means in layman terms no matter what ice or value you shoot as long as you stay in the native range of iso values you will achieve give or take the same results give or take that's important because there is no perfectly ISO invariant sensor but the Leica M11 is known to have that sensor and since my assumption is I said in the introduction to this video is that this PSI 60 megapixel sensor in the Leica Q3 is a close relative of the M11 sensor I assume that the iso invariants almost not perfectly is also given here in the new Leica Q3 here's another of these frames kind of the same story if you look into them left hand side High ISO 3200 right hand side 200 ISO corrected in Lightroom by four full stops on the exposure slider and these are essentially the same images so if we click into here there is not on the left or the right hand side more or less noise and these images are clear they are well represented here I think this looks good and we have the same conclusion it looks like an almost ISO invariant sensor in the same way as I explained before so it doesn't really play a role if we shoot these images with an ISO of 3200 on ISO of 200 the result will essentially again essentially is important here be the same and you can actually in your photographic parameters completely focus on image composition you know depth of field based on your aperture and of course on motion blurriness by making the shutter speed faster but you don't really have to bother about ISO whether you shoot them low or high eyes or doesn't play a role in post later on if you correct them these images are essentially the same now here's the last example of the same situation left hand side 3200 ISO right hand side 200 ISO corrected for exposure by four full stops in Lightroom and again these are essentially the same images now just by peeking and poking into these images and having your subjective view whether there is more noise on the left or the right hand side or if it is exactly the same is kind of speculative and I want to use a different approach which I didn't plan for when I started this video namely I will look deeply into signal to noise ratio Peak signal to noise ratio and also in particular on Pixel binning with the two lower resolutions the Leica Q3 offers now namely the 36 and 18 megapixel resolutions and I will try to provide a more substantiated approach what we can expect in terms of dynamic range from the new Leica Q3 and how we actually have to think about pixel binning technology which Leica says they introduced already for the Leica M11 and M11 monochrome and now found their way into the new Leica Q3 the results you have seen here they were all shot with 60 megapixels so full resolution but we have the option here on the Leica Q3 like we have it on the Leica M11 on dng resolution to also go from 60 megapixel to 36 megapixel and to 18 megapixel and the lower resolutions 36 and 18 they still use the full sensor area and that happens by pixel binning which is a technique used in many other cameras by the way also smartphones to increase and enhance dynamic range and create a better low light environment noise behavior and on the Leica Q3 this is implemented in the same way as on the Leica M11 and M11 Chrome and according to the Leica spec Sheets if you repeat the exercise I just showed with the lower resolutions you can actually count on a dynamic range of not 14 stops but 15 stops so the exercise will work even better so far when we talked about dynamic range we substituted it by the procedure to significantly underexpose then correct for the exposure in Lightroom and then getting about the same images with about the same representation of noise but that is of course a very blunt view to look at dynamic range so in order to get deeper into the topic let's talk about noise and before we talk about noise let's quickly talk about bits so let's assume we have an 8-bit grayscale image and jpeg for instance is a format which always comes with 8 Bits 8-bit means that the lowest value is a zero and the highest value is 255. a pixel value of 0 stands for perfect black a pixel value of 255 stands for perfect white let's say an image has resolution of let's say 10 24 times 1024 pixel then the image is represented by a matrix of 10 24 times 1024 cells and each cell has a value between 0 and 255 and you see here in green an example how this could look like we have all kinds of value but the lowest value we have in the cells of the Matrix is a zero and the highest value we have in the cells of the Matrix is a 255. let's make an example to get our hands on the topic of noise so for the following charts I used a software called pixinsight which is a highly professional post-processing software and environment for astrophotographers and on the left hand side you see I generated an image which has zero noise this is a perfect signal and to make my life easier I'm not dealing with values between 0 and 255 0 for perfect black and 255 for perfect white but I normalized it to a scale between 0 and one and I gave this image an intensity of 0.5 and it has absolutely no deviations in any pixel values at all from 0.5 that's why the standard deviation is zero and the minimum value is 0.5 and the maximum value is 0.5 and this is what you could call a perfect signal as you see in the histogram on the right hand side it just Peaks at an intensity of 0.5 and there's no deviation to the left or the right hand side now a standard technique in image processing and Signal analysis is to put the signal in perspective to noise and a very simple formula very much simplified if you want to use it here is we take the signal and divide it by the noise now our signal is perfect means absolutely the same namely 0.5 intensity on each and every pixel and if you want to divide this by the standard deviation there is no deviation as you see in the peak in the histogram on the right hand side so we would have to divide by zero which is is not allowed in classical calculus but if we approach 0 in the denominator the signal to noise ratio here can be called infinity or arbitrarily large let's now introduce Some Noise to the before Perfect Image where we had a perfect signal and you see in here on the left hand side the image has now some fine grain and I applied Some gaussian Noise here with a standard deviation of 0.07 which you can trigger as a parameter in picks inside which is called amplitude so 0.07 is our standard deviation and then the numbers statistically change because we have now a random distribution in a two-dimensional image here and we have still a mean of about 0.5 a standard deviation of 0.07 and the minimum value so the blackest black we got is 0.16 and the widest white we have in that image with that fine grain applied on a random basis is 0.85 so there's still no perfect white or black in the image but a lot of gray noise which Place between these numbers with respective intensities and if we now calculate the signal to noise ratio in the same simplified way as we did before we still have a mean of 0.5 but now we have a true deviation to the left and right hand side as you can see in the histogram here on the right hand side and the standard deviation is 0.07 and if we divide or build a ratio here we end up at the signal to noise ratio of 7.14 let's now increase the noise on the image on the left hand side by increasing our amplitude from 0.07 to 0.15 and you immediately see how much noisier and more grainy the image on the left hand side becomes by increasing the amplitude now in terms of statistics we still have a mean of about 0.5 and the standard deviation follows the amplitude so it's also about 0.15 and on the minimum value we now get very close to a perfect black and on the maximum value very close to a perfect wipe and you also see on the distributions or the histogram on the right hand side side that the distribution is even more symmetrically stretched now to the left and right hand side based on an increase of the amplitude and if we now calculate the signal to noise ratio we still have to divide a mean of 0.5 by now a standard deviation of 0.15 which gives us a signal to noise ratio of 3.33 if a compare this all together now from the left hand side signal to noise ratio of infinity in the middle Some Noise supplied signal to noise ratio of 7.14 and on the right hand side a noisier image with a signal to noise ratio of 3.33 and putting it in perspective to where we came from and started with we have on the left hand side a perfect signal and a perfectly High signal to noise ratio if some noise is supplied the signal to noise ratio goes down and if there is heavy noise the signal to noise ratio becomes even lower and that is a measure you can use besides many others for the quality of an image but most importantly in dynamic range compare comparisons and we come to a more complex version in a moment and also to some procedures how to obtain it we can move away from subjective statements like these images are essentially the same and base are statements on a more fundamental view which is based on some true hard-coded Matrix my little Excursion a moment ago into noise and signal to noise ratio was a much simplified approach for educational purposes only to help people not familiar with the concept to get what we are talking about here but if you want to go for a non-simplified scientific approach it becomes much more complex and here's an example of automatic noise estimation from multi-resolution support and that means you express signal to noise ratio in decibel you go for the mean of the squares of the signal you divide it by the variance of the noise you take it inside a logarithmic function to the base of 10 and multiply it by 10. the variance of noise then in the assumptions of this approach very often assumes a gaussian noise distribution with the zero mean and the multi-resolution support noise assessment measures then pixel variants in order to get an impression about noise in smoother areas of the image so there is some image processing going on and as photographers we know if we give up the iso value and digital cameras the noise is more elevated in the soft blurry background bouquet and in Darker areas in the shadows of the image and these areas are detected by techniques sometimes based on wavelets which is a technique going back quite a few decades already or in AI algorithms or by Deep learning networks there are different ways how you can identify the local areas in an image where you want to apply noise estimates in order to come up with the signal to noise ratio expressed in decibel here now two images I shot in the studio and analyzed in pics inside both from the Leica Q3 so here's the second image by the way just skipping forward and I did shoot them in raw but converted them into jpeg already and loaded the jpeg to pigs inside and the first one has 60 megapixel and 3200 ISO and the second one is also 60 megapixel 200 ISO same story as what we saw on the main train station in Zurich and I pushed it up four stops in Lightroom classic so that's the two images the only difference of these images if you look at them I skip fourth and back between them for a while are the background LEDs on the wall which are Dynamic and changed color to some extent and I should have actually switched them off but it will not move the needle on what I'm going to explain in the next couple of minutes I analyzed these two images in pics inside and use the embedded statistics which you can call and get a box here like you see in the upper right hand side corner and I also looked at the histogram of these two images and if we put them side by side if you look at these two histograms left hand side the high is O right hand side the low ISO but pushed up by four stops in Lightroom there again essentially the same I use that phrase again you see a bit of nuances in differences in the different color representations here but in general they look almost identical and the same applies to the statistics so in the statistics in picks inside you get the mean value of the pixels the mean of squares the variance the standard deviation the minimum the maximum and you can of course also check more boxes here and get a more enriched statistics but if you compare these numbers side by side I did this in a very diligent way they're essentially the same there are tiny little differences but it's really negligible what you see here in terms of differences and if I move forward and apply now the single to noise ratio comparison then I used a script which is originated by hardwood borneman and you can load this into pix inside and that calculates your the signal to noise ratio statistics and you see here in the upper part on the 3200 ISO image we get signal to noise ratios around 30 and on the image with 200 ISO we get about the same numbers it's not that different so here we see a real and true metric for comparing these two images from a signal to noise ratio perspective now this topic really caught my attention so I wanted to have a clean image manually focused outside in the garden no background LEDs mounted on the wall which change color and I shut this image and on the left hand side you see again the high eyes or image and on the right hand side you see the low ISO image pushed up in Lightroom by four stops and if you look at these images there is no noticeable difference and that's a qualitative statement of course but even if you stare for 20 minutes at these images you will not get any smarter and if I crop in by 100 they again look exactly identical and that is of course interesting and now let's try to get this in a more objective way presented by looking into signal to noise and if we do that we first of all consult the marketing material of the Leica Q3 and they say here we have these three resolutions the large dng the medium dng and the small dng with 60 36 and 18 megapixel and they say in the marketing material regardless of the chosen resolution the Q3 always uses the full size of the sensor which is another way of saying it's based on Pixel bidding technology and they further say in the marketing material for the Leica Q3 that if you shoot in ldng you have 14 stops of dynamic range but if you switch to mdng or sdng you get one stop dynamic range more and end up at 15 stops they also say on mdnt that this is the best performance ratio and that is something of course we'll try to find out in the next minutes in the video if we get this in a relative way then if we compare 14 stops with 15 stops we have about 7.1 percent more dynamic range in the medium and small resolution raw files within pics inside I did now run a lot of statistics and Analysis and I used raw files out of the camera not touched in Lightroom or post-processing at all and I always compared ldng with mdng and sdng and within these different resolutions I compared the high ISO image with the low ISO image and as I said I did a lot of analysis but it all boiled down to the following going from the left hand side to the right hand side and starting on the left hand side with ldng out of camera I found when averaging over the three color channels RGB that at the highest all level I got 30 decibel signal to noise and at 200 is oh I got 30.4 decibels secret to noise but if I switch to the mdng out of camera it gave me 33 decibel on the high ISO image and even 33.9 decibel on the low ISO image that means on high eyes over 10 percent gain in signal to noise and on the low ISO 12.8 percent up looking at the sdng out of camera it gave me on the high ice or a gain of five percent on Signal to noise and on 200 ISO 9.5 percent on Signal to noise and if we compare this now back to Leica specification where an increase from 14 to 15 stops when you go from LDN G to M or sdng you get 7.1 percent more dynamic range then here we gain even more and the M resolution is by far the most Superior when it comes to signal to noise and in this way also to dynamic range we just concluded that the M resolution is the best resolution if it comes to dynamic range with still a reasonably high resolution for cropping in 36.5 megapixel is plenty of space but let's look at this now from a slightly different but still related angle and let's look at this from the perspective of the difference image of two images and the way to measure the difference between two images is to go for the mean square error and you see here formula and in this formula m is the number of pixels in horizontal Direction n is the number of pixels in vertical direction i1 is image number one I2 is image number two and we compare the two in the formula both images are supposed to be converted into grayscale so instead of three color channels RGB we have only one channel and we only have light intensity here in these two images and have converted them before we compare them comparing two images and coming up with the mean square error allows us to calculate the so-called p signal to noise ratio in decibel and that's again 10 times the logarithm of 10 and in the numerator we have the maximum pixel value we can attain I explained earlier in the video that on 8-bit images this is 255 like in the typical jpeg image so we have 255 Square in Brackets in the argument of the logarithm and we divide 255 Square by the mean square error and we can also Express this as 20 times the logarithm to the base of 10 if we get the square inside the brackets out because in the logarithm if you take the square out it becomes 2 times 10 which is 20 and that's another representation you find very often in the internet and this is a typical representation for jpeg image and Peak signal to noise ratio will enable us now to in very precise manner compare High ISO low ISO high resolution low resolution images for the Leica Q3 before coming to the real stuff and going back to Leica Q3 images let's get familiar with the concept of mean square error and what it all means and in order to illustrate this I've written here a script in Python which reads two images from the SSD of my MacBook converts them into grayscale calculates the dimensions in terms of pixel 8 pixel width and then calculates the difference Matrix or a difference image between these two images and prints the mean square error to the console and the most simple example you can have here is an image 1 and image 2 as shown here on the right hand side so image one is a 1024 times 1024 jpeg image 8-bit grayscale all pixels have a value of 0 so represent the perfect black the second image same pixel Dimension 8-bit grayscale all pixels are represented by a value of 255 which is the perfect white and if I now read this into python I can print out these images as pixel Matrix and the perfect black Matrix which is the perfect black square is filled with zeros only the perfect white square is filled with 255 only because that's the value for a perfect white in these type of images with 8-bit and the difference Matrix then also contains 255 in all cells of the 10 24 times 1024 pixel Matrix or as we call it from now on the difference image the python script also calculates here a mean square error of 65025 and we see this in the formula easily because the 255 is the difference we have at the pixel level it's squared we can take it in front of the double sum and the ratio in front of the double sum and then it's 255 Square which equals 65025 and this is obviously if you look at the formula independent of the pixel resolution and it is the maximum MSE an 8-bit image can attain 255 square if on the flip side of the coin two images 100 identical then the mean square error of the difference image of the two images obviously is zero because at the pixel level every pixel difference will be zero if we now translate this into Peak signal to noise ratio in decibel the formula collapses to zero for our extreme example because it's the most extreme example you can have if you want to look at a difference image if you come from a perfect black square and look at its difference to a perfect white square and if you look at the formula it collapses to zero in our example and if in the opposite direction we would have two images that are 100 identical the mean Square Arrow of the difference image as we just saw would be zero and then the peak sequence noise ratio in decibel would basically go to Infinity you could also say it's not defined because you're not allowed to divide by zero but it's the perfect Peak signal to noise ratio in decibel if two images are 100 identical so I think I'm done with my educational session let's go back to Leica Q3 images and I have two of them here on display and one on our calculate the difference image of these two images again we code this in Python so we get to image 1 and image 2 on the SSD of my MacBook Pro we calculate the pixel dimensions we convert them into grayscale and then we calculate the difference image and we also calculate the mean square error and the peak signal to noise ratio if we do that first of all this is the grayscale conversion of the two images and that looks nice you see how easy that is in Python you just have a one-liner for each image super easy and then we have here the difference image and that looks quite interesting of course the difference image of two totally different images looks very messy but you see nevertheless in the difference image here nicely still the Contours of the cat as well as the structural elements from the architectural shot and the question is in the difference image which parts get a brighter intensity and which parts get a darker intensity in order to see this I've chosen here in the two images two areas and they are correspondingly the same on the two images and you see on the cap image these areas are darker and on the architectural shot these areas are very bright and if we now look into the difference image then we see these are exactly the two areas where the difference image is bright because the difference is larger largest difference as we saw in the extreme example of a black and white square is 255 but because the difference is larger the intensity is higher and it goes more towards some gray or white and that's what you can see in the difference image here so that is I think quite nice but the python script did not only give us that difference image in a graphic representation here we also can calculate the mean square error which is super high here 6628 and the peak signal to noise ratio which is low at 9.92 and the reason is that these two images are different that's why I have chosen this example what happens now if we turn our attention back to our high ISO and low ISO images which represent the same scene and here we see now the mean square error is only 129.611 compared to 6628 here on the different images and the peak signal to noise ratio is 27 almost exactly and we are here in ldng if we do the same for MD and G we end up at a mean square error of 68.55 and the peak signal to noise ratio of 29.77 and if we do this for sdng we get an MSE of 90.46 and a psnr of 28.57 so summarizing again we find here that if we go and calculate the difference image between the high is O and the low ISO image and then calculate for the difference image the mean square error and the peak signal to noise we get best results for the mdng resolution which is kind of in line of what we saw with our experiments within pics inside before let's quickly go back to the cat and the architectural shot and you saw here the mean square error was 6628 if we go back to the mdng and the two identical scenes but with high and low ISO value we got a much smaller value on the mean square error of 68.55 so if you put them side by side you see that the mean square error of the difference image of two times the same scene bus shut with different parameters is about in the order of magnitude of one percent only of the mean square error of the difference image of two images which by the way have a lot of similar bright and dark areas but are nevertheless totally different so that gives you a bit of an impression and a gut feeling on the order of magnitude of what you can expect here if you work with MSE of different images now with our python script we can not only calculate the MSE of the difference image and the peak signal to noise ratio as you saw before we can also visually represent the difference image and that's what I did of course also now with the right hand side images the high is all low eyes or representation of the same scene and if I do this here if I start with ldng you can stare at that image for a long time and on YouTube it will be hopeless anyway because of video compression you hardly see anything and so I took this image which is the native difference image and pushed it up on the exposure slide and Lightroom by three stops and then I see some structure coming up here on the mdng it looks almost the same but there is some artifact in the difference here in that red rectangle and if I push it up I see this even more elevated and emphasized here and the reason is very likely if I look into the original image where this part of the image sits where the difference Matrix has quite some difference and brightness and intensity it's very light likely that these leaves on the plants here were moving a little bit in the wind and you see this if you look at it and that's why we have here an elevated part where the difference Matrix shows some intensity which we not see in other parts of the difference image and then for sdng absolutely the same story the difference image as it comes natively out of the Python script almost pitch black if I push it up in Lightroom I finally see some structure and that's basically what we have and if I put these three different images for ldng mdng sdng side by side and stare at them despite the artifact we have on mdng because the wind was moving the plants a little bit which created in a local area some higher intensity I could swear that the middle one is the darkest one which means on average is the best because it has the lowest difference in terms of mean square error and I can of course also make this objective by again working with python and calculating for each of these three images the average pixel value and if I do that from the left hand side to the right hand side ldng has an average pixel value of 8.8 MD and she has an average pixel value of 6.2 and sdng has an average pixel value of 7.3 so again the average pixel value which stands for the average intensity is indeed the lowest on the middle image which is mdng and is confirming my visual impression when I look at these three images and if you put this in perspective and context the highest intensity we can have in 8-bit I've set this now over and over again is perfect white which is 255 so on LD and G we have on average 3.45 brightness on mdng it's 2.47 and on sdng it's 2.87 so that means in relative terms mdng is 28 better than ldng and sdng is 17 better than ldng if you want to look at the ranking of these three resolutions from this particular angle in Empirical research which is what we are doing here when we try to identify pixel pinning and dynamic range of the Leica Q3 it's always good to repeat an experiment with a different sample and in the one we just discussed the mdng had because of the wind between the high ISO and low ice or shots was blowing and these leaves here were moving a little bit we had these artifacts in the mdng difference image and so I wanted to repeat this in a clean setup in the studio and do the same experiment again and I also want to introduce a different metric to judge about the difference between two images and that's the structural similarity index whereas the mean squared error measures the differences at individual pixel level and by the way for identifying which of the two images the high or the low ISO image and the low one pushed up by four stops has more or less noise this is a really good measure but human perception if we look at images that does not happen at the pixel level humans look at an image holistically with respect to certain areas in the image for instance they look into Shadow areas and then identify some noise or into a blurry background and identify Some Noise there and the structural similarity index ssim assesses groups of pixels instead of working at the pixel level and trying to identify the intensity of the differences and it takes luminance contrast and structural information between two images into account and the two extreme cases we can have here is if for two images the ssim is one then these images are perfectly similar and if for two images the ssim is zero then these images are not similar at all let's look at a simple and well-known example from the scikit homepage I come to scikit in a moment in order to make that point clear so on the left hand side we have the original image of the cameraman here and if we subtract the original image from the original image we look at the difference of two identical images and then as we discussed before the mean square error is zero and the structural similarity index is one in the middle image we have the same image of the cameraman but with some noise and here now the mid square area is 0.04 and the structural similarity index is 0.50 on the right hand side image we have just increased the pixel intensity at each pixel by a certain amount and then the image is brighter and the mean square error now of the middle image and the right hand side image is the same namely 0.04 but the structural similarity index on the right hand side is much higher at 0.85 or you could say 85 and that means we have here two images they have exactly the same mean square error but look completely different when it comes to noise and the structural similarity index picks up on that because it recognizes in the middle image that there is a substantial amount of grain and noise and so the ssim goes down to 0.15 whereas on the right hand side image this is kind of noise free you have a structural similarity index of 0.85 so I went to the studio and shot a clean set of six sample images six because we have again High ISO and low ISO images of the same scene and then I did shoot them in 3D G resolutions and you see here the ldng left hand side High ISO right hand side low ISO and again if you stare at these images there is no noticeable difference we can do this for 20 minutes it will not make us smarter and even if we crop into 100 these images look kind of absolutely identical in terms of colors noise sharpness what have you it looks like the same image but I wanted to get this now calculated in the same way as before but in addition I want to include the structural similarity index and I don't repeat the whole procedure again because I did this over and over again in this video what you see here is for ldng mdng and sdng the three different difference images and I push them already up in light one by three stops if I would not have done it these three frames would be pitch black that would be absolutely nothing recognizable that's why I push them up so you see some structure here I calculated all the metrics we had before mean square error pixel signal to noise and an additional also the structural similarity index and I want to Now quickly walk through these numbers to make my points by the way in order to calculate the structural similarity index I used a library which I included in Python and you see here this is from scikit which is a large Library typically used in the context of artificial intelligence and machine learning with a lot of statistical algorithms and from the home page of scikit I also borrowed this image here which I discussed earlier in the video again the mean square error is by far lowest on mdng with 42.78 on Peak signal to noise ratio we find that the highest is on mdng should by now not be a surprise and again if I look at these three images I could swear that the middle one is on average the darkest so mdng again and if you do the same exercise as before and calculate in Python the average pixel value again it's confirmed mdng is on average the one we which is the darkest if you look at the structural similarity index we have a substantial uptick On mdng and sdng Here the ssim is substantially higher than what we have on ldng which again is in line with the Leica marketing material that there is one stop more dynamic range if you shoot in mdng or sdng and by the way if you look at the mean square error in percentage if you go back in the video I said that the mean square error in its maximum value can attain 65 025 so if you take the mean square error of ldng for instance which is 85.5 divided by the maximum mean square error we can have 65025 we end up at 0.13 percent only so also here we have a pretty good result if MSE is our metric it gets much better on mdng with 0.07 and still better than on ldng if we go to sdng with 0.09 so all in we see again that MD and G is by faster period there are always two buckets here there is the ldng resolution and there is M and sdng resolution and these two are always superior when it comes to the metrics we discussed in this video and in general two dynamic range I want to quickly conclude on the first topic of this video and don't worry the other two topics like insanely High ISO and how usable the images are and handheld shooting via Optical image stabilization these two topics will be very short so first of all the Leica Q3 sensor is essentially ISO invariant we see it by human perception when we look at images shot at high ISO and low ISO but the same scene and compare them and look at the level of noise we have in the image which we pushed up in Lightroom but we also see it by hard-coded metrics as I try to point out in this video there is true evidence based on the metrics we looked at so mean square error signal to noise ratio Peak signal to noise ratio and last but not least the structural similarity index that on the Q3 pixel binning is truly effective the dngm and dngs resolutions truly improve dynamic range and will be better in low light and it's confirmed that Leica says in the two lower resolutions you have one stop more of dynamic range I think there is enough proof of evidence in this video to can safely conclude that this is true and is implemented in an effective way and then the last conclusion is that mdng performs best and has still with 36 megapixel a pretty high resolution is about the resolution of an m10r so you have plenty of reserves if you want to crop in there is one important caveat which I don't have here on the slide namely that ISO invariants only holds within the native ISO range and the base ISO of the Leica Q3 is 100 and at a certain higher ISO value it will stop to show ISO invariance and when we in a moment look into insanely High ISO shots like 50 000 ISO or 100 000 eyes so then ISO invariants will no longer hold because it's only valid and working in the native ISO range everything I said on topic 1 is fully confirmed by the Leica marketing material if we look at what they say here mdng they say here in the red rectangle this is the best performance ratio it's exactly what I found in all these tests where I looked at the different resolutions from a different angle when it comes to different metrics how to measure performance and we looked into the different images of high ISO versus low ISO and in addition you also have of course a significantly lower file size on mdng namely 50 megabyte compared to 85 megabyte if you work with LD and cheese so all in shooting in mdng is always the best choice except you really need 60 megapixel for a high resolution print you cannot live for various reasons with 36 megapixels let's now move on to the second topic which is shooting at high ISO as I said I will do this in a couple of minutes show you a few sample images it won't take long and then let's look into ois based handhold shooting and what I got out of the do Leica Q3 without a tripod in the main train station in Zurich the second topic of this video is shooting with high ISO values and the iso range here on the Leica Q3 has been extended so the lowest ISO value is 50 but 50 is not the base ISO the base eyes are according to Leica is 100 now what does base ISO mean and in order to keep it simple let's call it the following if you shoot with an ISO of 100 you have zero amplification of any signals coming with the image and that should create the best possible image quality and then in contrast to the Leica Q2 where the highest ISO value was 50 000 we have now here an extended ice Orange going up to 100 000 and I will show you sample images which I shot at the highest eyes over here of one hundred thousand and to my big surprise they are still usable and of course I use some post processing in particular in Lightroom where Adobe included now ai based enhanced noise reduction algorithms but these images look really good and I should say with the exception of the monochrome cameras from Leica this is the first camera with a color sensor where an image shot with an ISO 100 000 in my experience is actually usable and that is of course a fantastic Insight which you can leverage in extreme light conditions wherever you are with your camera so here's the first sample image shot with an ISO of 12500 I had to go pretty fast on the shutter speed at this ISO level in the main train station in Zurich it's an ledng image so 60 megapixel and it's shot at an aperture of f 5.6 which typically on Leica lenses is The Sweet Spot and if you look at that there is substantial noise in the image already you see this here this guy is walking here but there is not a lack of detail look at the clock here in the background there is still plenty of detail in that image and now we have in Lightroom this AI based magic button here and the noise reduction which is called denoise so let's click on that and then it starts to load enhanced data and now we can see here the magic happening let's look at a different part in the image here maybe here let me go here to something where we have some writing and if I press and hold you see the amount of noise and here you see how it is hexed away so let's go for enhanced or let's play a little bit maybe with the amount let's go here to only 30. let's see if this is what you see there's already a lot of noise gone here in that image but let's go to 40 here then it looks even better and now let's go to enhance and this will take about a few seconds to be processed and then we are done with our noise reduction there is no other post processing applied here you see that all sliders are untouched and it's just the noise reduction which is going on here now it's done and you see the noise is now grayed out I cannot push it again and if you look here in the background now let's go here you see this is pretty clear you see all the writings here on these signs you see also the people pretty clear this is absolutely usable and clearly at 12 500 ISO is not insanely high but looks really good let's go to a different place here lots of details and very well rendered and again looking at the clock here very sharp and a very good image absolutely usable for social media no problem at all but also for smaller prints let's move on to 25 000 ISO and let me see where we have it here is 25 000 same story this one here has already been denoised so I cannot show you the before setting here but if you look into that still a very good image look here the small dance the schulhaus dash whites that means the most modern School building in Switzerland that's on this red dot here written in a very small way and it's still fully recognizable and the level of noise is quite contained it looks good no issues with that image so that's 25 000 and by the way I spoke before when we talked at the end of the dynamic range session about ISO invariance I said ISO invariance only works in the native eyes orange and I would guess the native ice orange for the Leica Q3 is from the base ISO which is 100 up to 25 000 and after that I don't think that ISO invariance will work because for many cameras 25 000 is the hard stop for the native ISO range let's move on here let's see what else do we have here so this is now a 100 000 ISO already denoised here is the image which is not the noise let's have a look into that here there is a lot of noise and grain here clearly this is 100 000 ISO but still if we look at the clock I can read this is 10 minutes past five in the evening and still here on these Engravings on the science year academic Gateway and I still can read in the Red Dot here that's more dance to schulhaus dash whites the most modern School building in Switzerland so this is quite acceptable and if you look at it from a distance already acceptable let's have a look into denoise now and you see there's a lot of noise let's go here now to 50 percent and let's go also zoom out a little bit here so here we zoom out and then if you go in on the faces the AI is almost recovering what should be there but not completely these faces here are distorted to some extent but that is not Percy a problem for me in social media because you will not be able to crop in that much let's go to the clock here in the background look if I press and hold you see the level of color noise and grain that was in the image before and then look what Lightroom does with the denoise Magic based on artificial intelligence so this is quite acceptable still for social media let's process this quickly and let's have a look again at that image the process is completed let's zoom in again look here this is absolutely amazing we still can read in that red dot here that's what the answer schweitz and clearly the AI has it much easier to recover any writing based on recognition then recovering a face in a pixel range so if I go here to this person you see these faces they don't look good if you look at them from behind you still see a lot of detail but it's not good if you zoom in crop in and what have you but if you look at that from a distance and if you are in a situation where you need to work with an ISO of 100 000 this is a reasonable approach if you see that image on the smartphone screen in Instagram that is still a usable image and of course in general I would not recommend to shoot images Beyond an ISO of 25 000 as I said before that's very likely end of story for the native eyes orange but if you have to you still get an image out of the Leica Q3 which you will not get out of many other cameras this looks quite interesting and still quite well the third topic of this video is how long I can shoot hand holding with the camera based on image stabilization incorporated into the lens and the way you can access this in the menu is the following you go to page number three and the first menu entry here is Optical image stabilization so that means stabilization here happens in the lens and not in the camera body like for instance for the Leica SL2 and sl2s and here I can choose between Auto and then I can also choose off or on and I have this typically on auto but on a tripod I very often switch it off because then image stabilization is not required and I was testing this out again in the main train station in Zurich looking at the escalators and then trying to handhold shoot as long as the image became not blurry in my shooting and my borderline was with half a second so 0.5 seconds was the longest exposure time I could shoot hand holding and that is a very good result for optical image stabilization you know with in-body image stabilization by a sensor floating freely in a magnetic field you might go up to one or two seconds but by Optical image stabilization only getting to 0.5 seconds is a really good result and with one second exposure the image became blurry I'm going to show this in a second now of course this is a subjective test because different human beings have different capabilities to hold a camera calm and sturdy in their hands and in my case it's half a second in your case you might be able to do this a second long and someone else might only get it for 1 8 of a second let's start what happened if I crossed my borderline which was 0.5 seconds so this is an exposure with one second again shot at F 5.6 in the ldng resolution and I had to go below the base ISO here and did shoot this with a nice over 50. and you see it emotion blurriness on the escalators looks actually quite nice and but then you see on areas where we should not have blurriness the image is blurry and I was not able to get this shot done in order to achieve a sharp image with a one second exposure shot handhold without a tripod in contrast to that if I went to 0.5 seconds and I tested the several times it works you still see here all the other parameters are the same as 5.6 ISO 50 ldng so 60.3 megapixel and you see here nice the motion blurriness of the escalators also here on the other escalator but you see the rest of the image is very sharp this is a usable image no blurriness any longer and it looks really good so this is my borderline you have to try this out for yourself as I said before I always do this when I get a new camera with stabilization to see what is my hard stop where I cannot shoot longer handhold than this particular exposure time if you like this video don't forget to drop me a thumbs up stay tuned on my channel there's always more to come thanks for watching stay safe and healthy and of course peace out

Channel: mathphotographer

Views: 14,722

Keywords: Leica Q3, Leica

Id: YH3LT4229ug

Length: 52min 55sec (3175 seconds)

Published: Sat Jun 10 2023