10 * 100 vs 100 * 10

[Frank7D]

"How exactly would a silent picture cause mechanical vibrations?"

It wouldn't; hence its advantage in astrophotography and the desire for longer possible exposures in FRSP.


Topic split from here.

[dmilligan]

I mean, why not just take shorter FRSP exposures and combine them? There's no vibrations to worry about.

[pulsar124]

That's what I'm doing right now, but apparently there is much more to astrophotography with DSLRs than the simple idea of "combine shots to increase S/N ratio". I discovered that effect myself, when I ran two stacking sequences of deep sky, one with 20 shots x 15s, the other one 100 shots x 15s, both at ISO 3200. I was puzzled to discover that both stacked photos had the same sensitivity (the same limiting magnitude), whereas I expected ~2.2x better S/N in the longer stack. Apparently this has to do with counting individual photons for these extremely faint objects, and averaging stacks no longer helps; one has to go for longer exposures to go deeper.

[dmilligan]

averaging stacks no longer helps; one has to go for longer exposures to go deeper.

I've heard this before and I'm not convinced that this statement is true (I'm also not convinced its false). I'd love to see some mathematical and experimental proof one way or the other.

There is some evidence that the bulb mode on the camera is actually taking multiple exposures and merging them digitally behind the scenes. If that's the case, I can't see how there could be any difference.

[pulsar124]

Well, I already provided one experimental proof: in my astrophotography tests, going from 20x to 100x 15s stacks doesn't improve sensitivity even a tiny bit (for Canon 50D at ISO 3200). Apparently ISO 3200 is a particularly bad choice (in terms of read noise; see http://sensorgen.info/CanonEOS-50D.html). I plan to redo my tests with other ISO, specifically ISO 1600 where the read noise is 6x lower. I am also testing different stacking algorithms (my original test results used the default Mean algorithm; I will test Median, Kappa-Sigma etc.)

[a1ex]

I'm not sure the exposures are combined digitally - the sensor readout is done once, at the end. But the sensor does appear to be reconfigured somehow periodically, and this happens at shorter exposures too.

Can you share the raw images for the 20x and 100x 15s stacks?

[Audionut]

If a photon from a faint object only arrives at the sensor every 15 seconds, and you are taking exposures every 10 seconds, then only every other image will contain these photons.  Stacking software if likely to consider these photons (wanted detail) as noise and discard it.

There's a difference between 100 * 10 seconds exposures, and 10 * 100 second exposures.  The later will capture fainter detail.  The former will only faithfully reproduce detail that delivers photons consistently every 10 seconds or less.  The total exposure time is the same, but they don't paint the same picture.


Nice work on the patch Greg.

[dmilligan]

Stacking software if likely to consider these photons (wanted detail) as noise and discard it.

Then that's a limitation of the stacking software not the methodology, and still doesn't provide me a convincing proof (a proof usually doesn't include the word "likely"). Even if you only see one photon every 100 exposures, as long as that single photon has a measurable effect, take enough exposures and average them and you'll see this signal. The only way you wouldn't is if there exists some sort of "bias", below which a signal has absolutely zero effect on the measurement (this bias could be a physical phenomenon, processing trickery in the camera, a result of rounding error in your calculations aka "the stacking software", etc.). Even though a single photon is going to be well below the noise floor of the DSLR electronics, it should still have some probabilistic effect on the measurement.

A definitive answer would include an explanation of what exactly is causing the "bias", and describe a controlled experiment that shows the existence of this bias. I think you guys are probably right, but I would like a convincing proof that explains exactly why.


(Sorry this is getting OT, feel free to split it into another thread.)

[pulsar124]

dmilligan: It's all down to pretty simple math (probabilty theory), no need to do too much experimenting. For exposures up to ~1min pixel noise is more or less constant, and it's well known that averaging measurements contaminated with gaussian noise improves signal to noise ratio as a square root of the number of measurements, so when you fix your total exposure (sum of individual stacks), the SNR will scale inversely proportionally to the number of shots. Say, 4x 60s will have two times better SNR than 16x 15s, with the same total exposure (4 min). But at exposures >1min thermal noise becomes important, and thermal noise scales with exposure time, so increasing exposure to >> 1min will no longer improve SNR, as both the signal and noise will grow at the same rate. (Actually, things are likely more complicated than that, this is just a simplistic outline.)

[dmilligan]

no need to do too much experimenting.

There's always a need to do experimenting. https://en.wikipedia.org/wiki/Scientific_method

(Actually, things are likely more complicated than that, this is just a simplistic outline.)

(a proof usually doesn't include the word "likely")

I'm asking for the actual complicated answer, not the simplistic outline.

[Audionut]

Even if you only see one photon every 100 exposures, as long as that single photon has a measurable effect, take enough exposures and average them and you'll see this signal.

Yes Sir.

But for that particular photon, SNR doesn't increase at the same rate as the other detail (when stacking), because..................it's only in every other exposure, not every exposure.  I'm pretty confident you don't put your lens cover on (for every other exposure) when taking exposures in bright conditions for reasons that don't require scientific evidence.