Puzzled by seemingly impossible improvements in SNR and DR over Sony BSI full-frame sensor, DxO checks Row files for pixel noise correlation, a sure sign of applied spatial noise reduction. However, there was no any correlation:
Then, DxO suspects that Red has applied a temporal noise reduction: "This technique, called temporal noise reduction (TNR), is most commonly used in video, since there are many successive frames to work with. However, temporal correlations across a time axis are not relevant when analyzing the image quality of a single RAW image, as they do not impact any RAW converters.
Whatever noise reduction system RED employs creating the RAW images from the Helium sensor, its presence means that we aren’t measuring just the RED sensor, so its results aren’t directly comparable to those from camera sensors we have tested from other vendors, whose RAW results come straight from the sensor with no prior noise reduction processing."
I have always wondered who REDs image sensor supplier is but they do a good job of keeping that quiet. Anynody know? Teledyne/Dalsa?
ReplyDeleteForza Silicon.
DeleteThis explanation of shot noise by DxO is quite unorthodox...
ReplyDelete"There is a certain inherent unpredictability in the behavior of photons. It is impossible to predict with certainty where a particular photon will land on a sensor. Depending on how large individual pixels are, some number of photons will land on a pixel other than the one you’d expect if they simply traveled in a straight line. This results in a certain minimum amount of noise (sometimes referred to as “photon shot noise”) when an image is captured."
The other side of the coin : they can benefit from a good training in the field ! See www.harvestimaging.com
DeleteWho are those other camera sensors who's results come straight from the sensor wit no prior noise reduction? I don't think Sony qualifies and since Nikon gets better scores with the same sensor I would think thye're doing some noise reduction as well. Pentax certainly is. Olympus and Panasonic do some too so maybe only Canon (their preference for Astro pics probably has something to do with this) but even then I wouldn't bet the house on it.
ReplyDeleteThere is also a fair chance that the designers of this sensor are better at CMOS image sensor design than Sony designers or Nikon designers. Performance is a combination of technology and design, everything else being equal. DxO historically has been weak on image sensor science and technology, as I recall.
ReplyDeleteEric, are you insisting that DxO has done the design of the sensor ? If so, then I think you are not right, DxO is not a design house to my knowledge. Their primary business is software algorithms I think.
DeleteNo, those were two separate thoughts. One regarding performance of image sensors and one a side comment about DxO in the past.
DeleteYes. Sure. But quite not fair, when comparing SONIKON and now even the Helium sensor. Canon for the first time in 5D4 presented CMOS with amps and ADCs all on the sensor, and since Dx0 measures only the "sensor (without the image processor) performance - Canon was always meant to be last before this last DSRL model in the CMOS performance competition. :/
DeleteIsn't CMOSIS the vendor of this 8K sensor?
Since Guy Meynants (CMOSIS) is asking about the SNR vs grayscale plot. I would assume not.
DeleteIt would be interesting to see the SNR vs. grayscale plot that DXO also measures. Maybe the answer is in there. The RED is not yet in DXO's public database.
ReplyDeleteIf it is due to to better CMOS design how do you explain the +6dB increase in SNR over leading FF sensor such as the A7Rii? There is no room for such gain with QE unless the colour separation is compromised, a non-Bayer array is used, or most probable..there is some yet undetected NR
ReplyDeletehttps://www.dxomark.com/var/ezwebin_site/storage/images/media/images/red_helium_weapon_8k__sony_a7r_ii__snr_print__920/156285-1-eng-US/RED_Helium_Weapon_8K__Sony_a7R_II__SNR_Print__920.png
Temporal noise reduction can be easily defeated by flickering objects. You can devise an object that switches from one gray-level to another fast enough. A single frame captured under such condition should not be touched by temporal NR. Otherwise, some intermediate level is seen and it means temporal NR worked in some wrong way.
ReplyDeleteYes, i have used a small fan with painted white blades to see the TNR effect. The blur of the blades with TNR enabled will appear more choppy than with it off.
DeleteThe RED is never in the full performance database, which is fishy IMHO.
ReplyDeleteAlso, the RED competition is not tested by DXO. What if Red is two stops better than a $400 usd still camera in video mode. How does it compare to ARRI? :)