12 replies to this topic

[Jeremy Batterson]

DEEPNIGHT AI NIGHTVISION--THE NEW KIDS ON THE BLOCK?

 

If the claims hold up that this new digital system can turn an ordinary CMOS sensor into a night vision system on par or better than Generation-3 analogue night vision intensification tubes, that will, indeed, be a game changer. In addition to the US military, the well-known digital low light sensor company Sionyx is backing this horse. Even if the claims do not hold up entirely, this will still be something of great importance for the NV Astronomical community.

 

P.S. The video link is not an ad, since the product is not yet for sale, so I think it is okay to post if for information only. 

https://www.youtube....h?v=keq2ewcXpIc

[TOMDEY]

That's gota be baloney.    Tom

[RichA]

That's gota be baloney.    Tom

Reminds me of the ads showing some guy putting on yellow-lensed "night driving glasses" and all of a sudden, the illumination level jumps 3 times.  

[GolgafrinchanB]

My initial impression is "plausible" but that they aren't really comparing the videos very fairly.

 

The "before" videos (on their website) IMO should be boosting the brightness to grainy levels but without their algorithm and the "after" video should be with their algorithm. Instead, they are presenting "before" videos without tuning the non-AI options way up from what I can tell.

 

The video in their lab setting also seems real. There's a ghosting quality to the motion indicative to me of some sort of statistical (AI) de-noising. I can see how it looks fake though since the clear shadows from their dim light source and WAY-TOO-DIM "before" settings produce something that looks like turning a light on. Also, the videos on their home page - while impressive - feature a decent amount of flickering and some other artifacts. I think they have more to go before I can use something like this without some noxious effects.

 

tl;dr - this looks to me like a form of LLM enhanced noise suppression in high gain video. I can see artifacts already and have my doubts that the views today are as pleasing to look at as something precise like night vision, but it does seem like a potentially cool technology. I'll be curious to see a product if I get a chance since I live nearby.

 

edit to add - I'm also curious whether or not they are including telescope views in their training data. In my experience, many AI models perform very poorly if you use them in contexts unrelated to their training data. All that to say, I'm skeptical with no plans to commit any $$ but glad someone's trying this.

Edited by GolgafrinchanB, 17 May 2025 - 12:53 AM.

[RichA]

So how far have electron-multiplying CCDs gone?  Last time I checked, they work, but have low resolution.

[rmorein]

If real, the explanation probably has to do with this:

 

https://www.cloudyni...c#entry13996303

[rmorein]

So how far have electron-multiplying CCDs gone?  Last time I checked, they work, but have low resolution.

Last time I checked, they required cooling.

[WheezyGod]

Would be great if it becomes a better alternative in the future. I’m still hoping something like the SPAD technology gets repurposed for live viewing and becomes less costly. I don’t see night vision being leaps and bounds better than it is now, but would love to be wrong.

The average NV device up until the recent L3 change was slightly better than what I saw on average 2-3yrs ago. Maybe something with a 37 SNR, and 0.5 EBI wasn’t overly difficult to find vs. 2-3 yrs ago you might only find these if you bought very soon after a vendor got a new shipment. Pre L3 change this kind of tube was a solid find but certainly not great or a unicorn tube.

5+ yrs ago, these specs were very hard to find. I remember reading threads suggesting a 33 SNR with a 0.5-1.0 EBI or a 28 SNR with a 0.1-0.2 EBI were solid tubes. The point is that if NV specs advanced a decent amount from 2020/2021 to 2022/2023, but only slightly from 2022/2023 vs. today, then the analog NV tech may have plateaued or be close to it. 10years from now maybe 40 SNR, 0.3-0.5 EBI tubes with 100k+ gain become the astronomical standard for NV devices, but we’d really need something with at least 55-60 SNR for the higher gains to be of more noticeable use.

Edit: I had called Steele who I bought my device from to ask about what my options are for warranty if I ever needed to use it with the L3 change and also asked them about this. Each military contract renewal tends to increase the specs which then requires someone like L3 to get tighter with their avg specs, but the guy wasn’t sure if something like a 60+ SNR was feasible if they spent more time of each tube or if this standard just isn’t feasible. He did reveal they got their highest ever SNR tube recently at 52.

Edited by WheezyGod, 20 May 2025 - 11:14 AM.

[PEterW]

Non-IT tech doesn’t follow moores
Law…. CMOS are close to perfect quantum efficiency. Bigger pixels and fast optics can help, this sort of computational cleverness can also help, but there will be physics limits never too far away. Digital needs decent amounts of power, but then so does thermal, so probably not quite as much of a problem as in the past when batteries were limited. Whether digital will eclipse analog for our sort of niche is an open question.

[TOMDEY]

Non-IT tech doesn’t follow moores
Law…. CMOS are close to perfect quantum efficiency. Bigger pixels and fast optics can help, this sort of computational cleverness can also help, but there will be physics limits never too far away. Digital needs decent amounts of power, but then so does thermal, so probably not quite as much of a problem as in the past when batteries were limited. Whether digital will eclipse analog for our sort of niche is an open question.

For what it's worth... analog is experiencing a huge comeback in developmental research technologies, especially in the arenas involving ~Artificial Intellegence~ to the point where many quantum computing players are predicting that will be the future after only a modest interlude centered on solid state digital. This also (creepy and exciting) correlates with my prediction from nearly 70 years ago that "computers will be grown" when we figure how to do that. At that point, sentients will once again become more bio than robotic creatures roaming the cosmos. And that may well explain where we came from, in the first place. We explorers are simply seeking our roots, claiming to be "scientists" and unaware that we are still/nevertheless/thankfully --- entirely primal.    Tom

[rmorein]

For what it's worth... analog is experiencing a huge comeback in developmental research technologies, especially in the arenas involving ~Artificial Intellegence~ to the point where many quantum computing players are predicting that will be the future after only a modest interlude centered on solid state digital. This also (creepy and exciting) correlates with my prediction from nearly 70 years ago that "computers will be grown" when we figure how to do that.

John von Neumann
 

"Hans Bethe on von Neumann

Nobel Laureate Hans Bethe said “I have sometimes wondered whether a brain like von Neumann’s does not indicate a species superior to that of man”, and later Bethe wrote that:“von Neumann’s brain indicated a new species, an evolution beyond man”.

[TOMDEY]

John von Neumann
 

"Hans Bethe on von Neumann

Nobel Laureate Hans Bethe said “I have sometimes wondered whether a brain like von Neumann’s does not indicate a species superior to that of man”, and later Bethe wrote that:“von Neumann’s brain indicated a new species, an evolution beyond man”.

It may have been staring us in the face all along... without our even realizing it... "the obvious"? It would be ironically amusing if ~the gods~ are real...and what we aspire to become.   Tom

[rmorein]

It smells of a neural network, maybe related to Hopfield, trained to differentiate the changing elements of a mostly static scene.

 

If so, it's great for security cameras and useless for astronomy.