Merged Discussion of femr's video data analysis

As an extreme example, does the acceleration pattern seen in the purple graph seem realistic to you?

[qimg]http://www.cesura17.net/~will/Ephemera/Sept11/Femr2/sg51.png[/qimg]

For an insect, maybe. For a large, heavy object, no.

Not all output can be considered to be representations of the actual acceleration. One has to experiment in order to move toward the most realistic solutions.


Yes.

Of course, I've worked with actual accelerometers and raw data recorded by accelerometers, which usually does look like the purple graph (because things that move also tend to shake and vibrate), so I don't have the benefit of vague intuition telling me what such data "should" look like.


I tend to agree with you: Yes, the actual acceleration profile of a precise point on the surface of the NW roof corner - say, a specific radioactively marked molecule - might be that wild, because the material it's sittiing on shakes and vibrates as the building rubles to a violent fall.

It is however, with utmost certainty, NOT the acceleration profile of the purple graph. That is because we can't measure that point with the sort of precision I indicated, and instead each individual measurement is fraught with several errors. The purple line probably is dominated by the random fraction of those errors, not the actual fine movement.


I'm going to pick up from here, if I may.

First of all, Oystein is correct in saying that the purple graph is not an accurate reconstruction of the actual acceleration. There is no way to reconstruct any actual high-frequency signal that was lost in between the video frames. (Fractal interpolation can generate a likeness of such a signal, if one knows or assumes it was there in the first place and makes certain assumptions about the underlying spectrum; thus, a fractally compressed photograph of a forest can be enlarged beyond the original size of the photograph, and will show interpolated detail that was not there in the original photograph, but that will not be the same detail that actually existed in the scene when the photograph was taken. It's quite likely that the purple trace is doing essentially that.)

But that wasn't the question I was answering, which was "does the purple trace seem realistic?" To which I answered in the affirmative.

The reason it seems realistic is the nature of short-timebase accelerations. Most people have reasonably correct intuitive understanding of sustained accelerations such as in vehicles or falling, and even an approximate recognition of their magnitudes (such as, half a g lateral acceleration is impressive performance for a road car, corresponding to a zero-sixty time of about five seconds). For faster events such as collisions, people's intuitive sense of the magnitude is poorer, which is why for instance public service messages have to remind people that a baby held in arms in a car cannot be held onto in a moderate-to-high-speed collision. And for very fast events, most people don't have the slightest clue.

For example, one accelerometer device I have used extensively is designed for recording human activity on the scale of hours to days. The accelerometer in it triggers whenever accelerations of 0.1g or greater occur. You can hold the device in your hand and move it around very smoothly and gently, and it will not trigger. But if you hold it in your hand while making normal gestures, or very gently shake it, it will trigger. That's all pretty intuitive so far.

Now suppose I rest the accelerometer on a desk or table -- such as a sturdy steel office desk with a heavy top, weighing a total of 200 pounds. Without touching the accelerometer device itself, what do you think I would have to do to the desk, to get the device to trigger? Perhaps, grabbing the edge and shaking with all my strength? Or lifting up one end a few inches and letting it slam back to the floor? Or hitting the desk with a sledge hammer?

Actually, all it takes to trigger the device is a gentle tap with one finger, with about the same force you'd use to type a character on a computer keyboard, anywhere on the desk top.

And that's not a flaw or failure or "measurement error" in the device. The finger tap does actually cause a >0.1g acceleration of the desk top to occur, so the device is correct to trigger as a result. The acceleration just occurs for a very short time, producing what we would feel as only a very slight (if perceptible at all) vibration. Our intuitive sense of the magnitude of that acceleration is way off, because the time and distance scales involved are so small.

And note that we're not talking about a molecule here as Oystein suggested, nor the insect posited by Major Tom. A big steel office desk.

Now suppose instead of a single desk, we use a heavy counter that's bolted to the wall and runs the entire length of the room. Does the finger tap still work? Yes it does, if it's within a few yards from the device -- because while the larger counter has more mass (and the counter + the wall it's bolted to + the rest of the building the counter is part of far more still), objects generally get less rigid as they get larger, so the finger tap can still vibrate the portion of the counter near the accelerometer, so as to cause >0.1g acceleration.

Now perhaps you can see why if you affix an accelerometer to, say, the dashboard of a car prototype performing a quarter mile acceleration test (aka a drag race), the data you get will not be a smooth trace of the car's acceleration from zero to eighty mph. It will be full of spikes and swings of a full g and more, much like Tom's purple trace, or likely far worse. And --

-- yes, here is where I get to the point --

that's not noise.

That's the actual acceleration of the dashboard.

(That's not to say that there might not be additional measurement error or instrument noise on top of that. Just that even without any such error or noise, you'll still get an incredibly messy-looking trace.)

You can only call some aspect of the data trace "noise" once you've defined what it is you intended to measure or evaluate.

For example, an engineer interested in the overall torque output of the prototype car's engine and drive train, who wants to know how well the car accelerated down the road, would use a low pass filter on the data, to filter out all the high-frequency signals, the clunks and rattles and vibrations that don't figure in "how fast this baby pulls out" and so are just noise for his purposes.

And in the adjacent office, another engineer interested in the performance of the vibration-damping engine mounts, would use a high pass filter on the same data, to measure how much high-frequency vibration from the engine was getting through to the dashboard at various speeds. She doesn't care about the down-the-road acceleration, or lower-frequency bumps in the road, so the part she filters out is just noise for her purposes.

And in other adjacent offices, there's an engineer studying the performance of the car's suspension, and another studying the smoothness of the clutch during gear shifts. They would use two different bandpass filters on the same data, to filter out the road acceleration and the engine vibration and everything else that's just noise for their purposes.

None of them, though, are getting the "true acceleration" out of the data. The closest thing they have, or could ever get, to the "true acceleration" is the raw data itself. (The only exception to that would be that if there were known sources of instrument noise, with known characteristics e.g. a known frequency and magnitude or a known DC offset, those could be compensated for or filtered out before the data is passed along to the four engineers.)


So, how does that relate to the measured movement of the top edge of the outer curtain wall of Building 7?

Barring any rocket engines (or helium balloons, winch cables, etc.) in operation at the time, the only known constant influence on the structure is gravity.

By using a very low-pass filter on the data (that is to say, smoothing), you can get a pretty curve showing the net result of the influence of gravity on the wall. But in that case what you're filtering out is "noise" only if you were trying to use the fall of the building to measure the acceleration of gravity. Which I think I can safely presume is not your objective here; there are many far better and easier methods for measuring g.

For every other purpose, what you're filtering out is part of the signal. Such factors as "resistance" of the lower structure, which would have to involve multiple separate instances of breaking and crushing of structural members; collisions and impacts in the course of collapse, including those that could transfer impulse from other hidden parts of the structure; and even the effects of explosives, if there were any; would all involve higher-frequency events that you're filtering out, and likely would involve magnitudes of acceleration both positive and negative of a significant fraction of a g and more.

Which leaves me, like others, at a loss for what purpose this can serve.

Respectfully,
Myriad
 
But that wasn't the question I was answering, which was "does the purple trace seem realistic?" To which I answered in the affirmative.
The level of noise is significant, reaching roughly +/- 0.2 pixels for the Dan Rather data.

That is measurement variance during time it is known that the entire building was not "jumping up and down", but instead was essentially static.

For example, one accelerometer device I have used extensively is designed for recording human activity on the scale of hours to days.
I don't really think it is valid to compare the obviously very sensitive accelerometer devices with acceleration data generated by extracting second order derivative from noisy displacement-time data.

Especially when using a very narrow window for determining average acceleration, which of course massively amplifies any noise in the data.

Given the scale of the building, and limitations in how accurate a position can be determined, I'd imagine their is in the real world a fair bit of damping of motion going on at the NW corner of WTC7.

that's not noise.
I agree that a certain amount of "wild" fluctuation is derivative acceleration will be "real", I suggest in this case it is swamped by actual noise amplification.

In my opinion that is an assertion supported by the fact that we already have multiple datasets (Dan Rather and Cam#3) which have different noise levels...whilst tracing (pretty much) the same point on the building.

It is very clear to me that, for example, tfk's suggestion that WTC7 was experiencing accelerations of +/- 80G, from derivatives using central difference approximation with the smallest window possible...are not "realistic".

Similarly, the extreme accelerations in WDC's graphs are also undoubtedly primarily the result of amplification of measurement error and also not "realistic"...an assertion I suggest is bolstered by the reduction in erratic acceleration behaviour during descent. There is "wilder" acceleration behaviour in WDC's plot when the building is not moving...

If there WAS an accelerometer on the NW corner from which actual data had been gathered things would be different, but there wasn't.

None of them, though, are getting the "true acceleration" out of the data.
Of course, thus my focus on trend.

NIST performed a linear regression on velocity data using 10 datapoints manually recorded over a 2.25s period.

I've revealed a clearer picture of acceleration trend behaviour by using high quality tracing facilities and extracting 60 samples per second, which of course requires application of averaging/smoothing to get useable information.

For every other purpose, what you're filtering out is part of the signal.
To some extent, of course. I think that it is clear that the various trend shapes are very indicative of the "true" trend behaviour.
 
Last edited:
Duh. I was trying to reproduce your results in order to make a point about your method

Yet another case of "prove femr2 wrong version the umpteenth" :rolleyes:
Actually, one of my points was to show the forum members that your acceleration graph did depict an acceleration curve, not souqtfmcA, and show what it was the actual derivative of (with "derivative" defined as "visual slope of the curve at each point").

Which can be summarized as "prove femr2 right version the umpteenth".


A particularly clear indication of "prove femr2 wrong version the umpteenth". Top job, er, no, wait...

:D
:rolleyes:
 
Actually, one of my points was to show the forum members that your acceleration graph did depict an acceleration curve, not souqtfmcA
Which is great, not that there was any real justification for the ridiculous term. All current curves, Poly10, Poly50, Symmetric Difference (13), SG (super-sekrit methods ;)), Sg 1 pass 2nd derivative (2, 51), All of WDC's various SG curves, tfk's dodgy decimated and interpolated curve fit, your own SG smoothed and symmetric difference derivatives, ... all show the same basic acceleration trend.

There was never any need to "show the forum" anything, but thanks ! :)

and show what it was the actual derivative of (with "derivative" defined as "visual slope of the curve at each point").
Have you managed to find a suitable SG implementation which generates derivatives ?

As I said, I'm glad you're delving into the data, but if the only purpose is to validate/trash a couple of femr2 graphs, what's the point ? You've generated your own acceleration trend data from my base data, for both Dan Rather and Cam#3. I'd call that something in the direction of your own "analysis", but it seems from you saying the particularly flawed "But I'm out of that game. If you don't want to publish your methods, your results are unverifiable and thus useless." you veer right back onto "prove femr2 wrong version...

So, you're not a fan of the 2 year old femr2 SG plot for acceleration. No worries. How about the 2 week old 1 pass SG plot for acceleration...
621711303.png


How bout that one ?


I note you haven't responded to my posts about early motion...

pgimeno,

No response about early motion yet...

The most direct visual "proof" of early motion is, imo...

[qimg]http://femr2.ucoz.com/_ph/7/787997325.png[/qimg]

To simplify...

a) Until t=60 static region motion and west edge motion are very closely matched = no motion of west edge

b) After t=60 west edge moves more than static region = motion of west edge

c) The difference between static region (camera) motion and west edge motion is the actual motion of the west edge


Significant motion of the West edge of WTC7 begins at t=60, which is ~2 minutes prior to release.


Do you have any issues with this pretty darn clear "justification" ?

Any response ?
 
...Which leaves me, like others, at a loss for what purpose this can serve...
I find it difficult to see how anyone can be "at a loss" to comprehend "what purpose this can serve" in the context of the topic of this thread.

But let me sort out some of the multiple ambiguities in the quoted part of your statement to support what I am saying.

First is the ambiguity of what you mean by "this" which is whether:
a) you mean "this" in the setting of the strawman and derail construct of the final paragraphs of your post; OR
b) do you mean "this", in the context of the topic of the current thread. In that meaning "this" is femr's measurements and derived data specifically for the current example of the collapse of the NW corner of WTC7.

So for my post I will take "this", in the context of the current discussion, to refer to femr's measurements and derived data. I will not fall for the derail to explain the flaws in the somewhat contrived strawman of your claim that "what you're filtering out is "noise" only if you were trying to use the fall of the building to measure the acceleration of gravity.". That statement is wrong even in its own context but...I'll avoid the derail. :o

Second issue is what do you mean by purpose? And, on this one, the ambiguity is not of your creation.

The reality is that there are two purposes - or rather one single purpose stated multiple times by femr and a group of related multiple purposes stated by other members.

And those purposes have been explained many times:
1) First is femr's own purpose in measuring displacements, velocities and accelerations against a time base and to a high accuracy. Prima facie he has done that. All attempts to prove him wrong posted to date present only partial arguments and therefore are not substantiated. I will show later in the post what I regard as the main error.
2) Second is the purpose of other members who seek to use the data. (Or group of other members purposes - several have actually been posted. Both legitimate and illegitimate. :)) We shouldn't lose track of the fact that on this occasion and several others femr has provided the data in response to requests from members.

So femr's purpose on those occasions at least was to meet members requests for data - their purpose was their purpose not femr's. Put more colloquially "It was their business what they used the data for."

Bottom line for this aspect of "purpose" being that followers of the topic should be well aware of the two purposes. The history of attacks directed at femr abounds with conflation and obfuscation of the clear distinction between his purpose and other members purposes. Which is part of the reason why I find it difficult to comprehend that anyone who has followed this debate and related topics can claim confusion over "purpose". Other than those who use pretended confusion as a deliberate ploy but I would not expect that from members currently active in this thread.

Now the "main error" that I referred to is quite simple and I have already tried to approach it several times.

The data we are discussing goes to measurements or derived values for d, v, and a against t.

It consists of the broad trends of d, v and a which are of interest for "the purpose" which are marred to some extent by sources of error in two main arenas viz "noise" and "procedural artefacts". And I use those two categories for purposes of explanation - not as targets for nit picking pedantry. If I have missed something or there is a strong desire to call it something else tell me. Provided it is not a ploy to defocus discussion I'm quite amenable to going along...

The issue of concern therefore is "are the noise and/or artefact induced errors big enough to affect the validity of the results for 'the purpose' to which the results are to be put?"

That question cannot be answered without reference to 'the purpose'. And the answer could well differ for different purposes.

AFAICS none of the claims opposing femr's work have so far attempted to show that the alleged errors could introduce errors of magnitude sufficient to invalidate femr's findings. And the shortcomings of argument go to two related aspects viz:
(i) None have quantified the scope of error to show that the error "swamps" the underlying finding of d, v, or a; AND
(ii) None have (again AFAICS) defined the purpose for which the claimants are seeking to invalidate femr's claims.

And I have not lost sight of the fact that one valid purpose is femr's own purpose of provide accurate data - to the standards of accuracy he claims.

(IMO) "they" in effect are claiming "here is a noise factor OR here is a mathematical process source of error THEREFORE since there is such an error the values are unusable".

Some of the claims seem to go as far as to say "...since there is such an error the values are totally wrong".

Then again there are some "no-man's land" issues of "I don't know yet?" - leave those for now.

So:
...Which leaves me, like others, at a loss for what purpose this can serve...
What "purpose" do you have in mind? femr's? One of those posted by other members? or yours from the referenced post which is, IMO, either not the topic of the thread or too loose a parallel to be a useful analogy?
 
Last edited:
Can you support that with citations?

:D Would imagine this from you only a few posts earlier would qualify...

If you don't want to publish your methods, your results are unverifiable and thus useless.

...especially as my results do not all relate to one single plot produced nearly two YEARS ago :eye-poppi

...especially as you'd been provided with a "1 pass SG" plot for acceleration, which you were provided full details to replicate...only a WEEK earlier :jaw-dropp

But hey, even though you ask about early motion results (which of course are extreacted from plots which have naff all to do with derivative generating techniques), and are provided details, you just ignore them...because you want to whine about the fact you can't replicate one graph out of a hundred.

Good job sir ! ;)

ETA: Needless to say, if one was to scrub back through this thread there is undoubtedly an abundent pile of assertions similar in form to the suggestion by ozeco41.

ETA: Needless to say that numerous acceleration profiles have been provided...Poly10, Poly50, "femr2 SG", 1-pass SG, Symmetric Difference (13)...your...tfk's...WDC's...all of which show the same basic trend...which is a significant improvement over those from NIST, Chandler, ...so that you say "[my] results are ... useless" says lots about "where y'r at" than anything else.

ETA: And, again, it's good to see others (you) generating your own plots. No-one said you had to use mine. You've come up with the same general trend too...fanks :) "validating" my "results". Splendid.
 
Last edited:
:D Would imagine this from you only a few posts earlier would qualify...
There are multiple examples some explicit, some implied.

I deliberately set out what should be a more challenging higher level argument. But pgimeno ignored the substance of the post in order to comment on a throw away minor point. His comment looked like parody so I let it pass. Didn't want to respond with laughter in case I was reading it wrong. :rolleyes:
 
Some of the claims seem to go as far as to say "...since there is such an error the values are totally wrong".
Can you support that with citations?
:D Would imagine this from you only a few posts earlier would qualify...

If you don't want to publish your methods, your results are unverifiable and thus useless.
No, it would qualify as a distortion of these words, which is why I asked, and you swallowed it. I am fed up of these kinds of distortions on Ozeco's side.
 
Last edited:
Please address the argument

I don't think pgimeno wants to.

From prior discussion...

femr2 said:
You are aware that any doubts you have about methods of generating derivatives have absolutely zero to do with identification of early motion, yes ?
Yes. But when such identification is made by the same person who makes such blatant mistakes, I have to doubt the identification, and the standard for proof is raised.
1) The "mistakes" pgimeno is referring to is simply that I have not disclosed the exact details by which my 2 year old acceleration plot was generated. Not "mistakes", simply "stuff pgimeno doesn't know".

2) pgimeno believes femr2 has made "mistakes"...therefore everything he says is in doubt, regardless of clarifications provided.

femr2 said:
So, are you in doubt about the ability of SynthEyes to detect sub-pixel scale positional changes in traced feature region ?
No.
...yet pgimeno still chooses to reject simple displacement/time data.

pgimeno,

You stated...
I challenge his explanation of how he achieved his results.
You were referring to Early Motion data.

You have been provided full details of how early motion conclusions were drawn.

No response.

Whether more early motion can be visible in the NW corner is what I am challenging at this stage.
Now that I have, again, reposted the details surrounding early motion, do you now "accept" the assertions about early motion ?

If not, why not ?

We're in a very simple logic case in essence: "femr2 methods imply early motion".

I am claiming femr2 methods are wrong.
You are quoted above saying that you agree that SynthEyes detects sub-pixel motion.

Assertions about early motion do not involve any kind of derivative, simply displacement/time data.

What "methods" do you have issue with ?
 
I don't think pgimeno wants to....
The phenomenon is quite common. Quite a lot of forum discussion has a focus on details. It can be sound discussion but can equally be off track in the full context.

Plenty of people are happy arguing at detail level so, as your know, I tend to look for the broader picture. Few want to go there so more often than not my comments get ignored. Leading to me spending countless sleepless nights fretting and worrying......er....not. :)

Whether they wont go there because they cannot handle reasoned discussion or they wont go there because it doesn't suit their tactics of debate is situation specific. And often includes both aspects.

There is potentially, in this thread, a very interesting discussion about the validity and accuracy of your work - whether stand alone or in comparison to NIST and others. But no-one wants to discuss it other than in the brilliant sunshine coming from a cloudless blue sky. ;)

I'm having fun on at least two other threads pressing for discussion to stay on topic - if "they" do the discussion is over. So I'm getting little support from either side who are into discussions at details level and have no concern that the details are irrelevant to the OP. Provided they can keep stirring each other up I suppose. :(

They needn't worry too much. I'll grow tired of the game soon. :D

...1) The "mistakes" pgimeno is referring to is simply that I have not disclosed the exact details by which my 2 year old acceleration plot was generated. Not "mistakes", simply "stuff pgimeno doesn't know"....
The confusion of objectives I tried to clarify recently.
...2) pgimeno believes femr2 has made "mistakes"...therefore everything he says is in doubt, regardless of clarifications provided...
The error of "false global" claims which another of our colleagues is fond of employing. "I find one bit wrong therefore everything is wrong." Irrespective of "relevance" or "significance" of the "detail" in the "bigger picture". :):rolleyes::o
 
Busy here. Patience.
Not much of that here, I see.

During this time I have read the Savitzky and Golay paper, and given a first quick reading to two other related papers that enhance on theirs, that were published in the same journal, namely:

  • Comments on the Savitzky-Golay Convolution Method for Least-Squares Fit Smoothing and Differentiation of Digital Data, by Hannibal H. Madden [link]
  • General least-squares smoothing and differentiation by the convolution (Savitzky-Golay) method, by Peter A. Gorry [link]
The first seems to point out errors in their tables. I have to reread it to remember what else it addresses. The second one establishes the bases for extending the SG method to produce smoothed values without discarding (W-1)/2 values at each end (W being the window width) and is also probably the method used to extend SG to asymmetrical windows.

One of the things that called my attention is this bit from the original SG paper:

Repeated Convolution. The process of convolution can be repeated if desired. For example, one might wish to further smooth a set of previously smoothed points, or to obtain the derivative only after the raw data has been smoothed. Thus, if we convolute using p points the first time, and m points the second,

[equations VIII, IX, and X snipped]

thus,

[equation XI snipped]

where [snip]

Equation XI shows that one need not go through the convolution procedure twice, but can do a single convolution, using 2(m+p)+1 points, and a table of new integers formed by combining the c's .
​

Interesting, isn't it? Maybe two SG passes are not equivalent to a single SG pass with different parameters, but they are definitely equivalent to a single convolution pass with different parameters. ETA: And note how the window width is 2(m+p)+1, therefore it's influenced by points (m+p) far!

I have also been able to run the Python SG function, which does produce derivatives, after adapting it (my Python was too old to run the function as is), but I still need to make a proper analysis of the derivative and compare it to central differencing. This is my program and my (very minor) changes:

Code:
import numpy as np
import sys

def factorial(x):
    p = 1
    while x > 1:
        p = p * x
        x = x - 1
    return p

def savitzky_golay(y, window_size, order, deriv=0, rate=1):
    [...]
# Comment these out as my 'math' module doesn't contain factorial, and numpy is already included above
[strike]    import numpy as np
    from math import factorial[/strike]
#    import numpy as np
#    from math import factorial
    [...]

y = np.loadtxt(sys.stdin)
y = savitzky_golay(y, 19, 1, 0)
#y = savitzky_golay(y, 31, 1, 1)
for i in y: print i
(I edit the file to run the filter with different parameters.)

I have also checked that the differentiation method used by QtiPlot consists exactly of averaging the slopes of the straight lines going from prior-to-current point and the current-to-next point. It should be obvious from the formula used in the source code, line 72 here: http://svn.berlios.de/viewvc/qtiplo...Differentiation.cpp?revision=1339&view=markup

AverageOfSlopes.png


It's easy to prove that for equispaced data (hope that's a proper English word, the spell checker marks it wrong but "equiespaciado" is a valid Spanish word), which is the case for the video data, that formula's result matches exactly the slope of the straight line that goes from the previous point to the next point.

It's hopefully obvious why that method produces visually correct derivatives consistent with the mathematical definition.

All that without having much time at all. I will continue and reply when I can.
 
Last edited:
Interesting, isn't it? Maybe two SG passes are not equivalent to a single SG pass with different parameters, but they are definitely equivalent to a single convolution pass with different parameters.
Correct.

ETA: And note how the window width is 2(m+p)+1, therefore it's influenced by points (m+p) far!
Correct.

I sympathize with those who haven't been paying close attention. Here's a brief summary: If femr2 obtained his alleged Savitzky-Golay acceleration curve by applying Savitzky-Golay smoothing twice, as appears to have been the case, then
  • femr2's method was equivalent to a single-convolution method, even though femr2 performed two separate convolutions around each point and has been pretending those two convolutions were necessary.
  • femr2's method was not equivalent to Savitzky-Golay differentiation.
  • Each point of femr2's alleged Savitzky-Golay acceleration curve was influenced by points within a window that was about twice as wide as femr2 has claimed.
(For the next few days, I will not be reading this subforum. I just wanted to point out that pgimeno is right about this, no matter how much technobabble may be posted in response.)
 
Not much of that here, I see.
If I didn't have sublime levels of patience, self-control and wasn't a thoroughly nice chap I'd have been out of here years ago.

Consider...

16th Feb...
He says there is early motion minutes before release

I responded 1 day later, providing you with at least four posts on the subject of early motion...

Here
Here
Here
Here

...following which you've made about 10 posts within this thread and said nothing about the information provided to you, that you effectively asked for (even though it was already within this thread, and I've reposted it simply to relieve you from the arduous task of running a quick search).

During this time I have read the Savitzky and Golay paper
What were you basing your understanding upon beforehand ?

the window width is 2(m+p)+1, therefore it's influenced by points (m+p) far!
We discussed that several weeks ago.

I still need to make a proper analysis of the derivative and compare it to central differencing.
Why ? Curiosity ?

It should be obvious from the formula used in the source code
1 sample width ! Ew :(
 
If femr2 obtained his alleged Savitzky-Golay acceleration curve by applying Savitzky-Golay smoothing twice, as appears to have been the case, then
Ha ha ha. Refocus upon one graph from two years ago that you're still trying to guess at. Priceless.

femr2's method was equivalent to a single-convolution method, even though femr2 performed two separate convolutions around each point and has been pretending those two convolutions were necessary.
Say wha' ? :rolleyes: Here's the thing William...it's not me that's stuck in a banal and anal "math masturbation" loop. I generated acceleration data from velocity data. Simple as. Your ludicrous assertions about "pretending" single/double convolution (in)equivalency are... just nonsense.

You still don't know how the graph was generated. Must be frustrating for you. Flapping about in the dark for two years. Great for the reputation that. Never mind eh :p

femr2's method was not equivalent to Savitzky-Golay differentiation.
If you mean a 1 pass 2nd order derivative from displacement data, then...duh...of course not...

Here they are again...
621711303.png

Hint: You might spot that the two SG plots are not the same ! Astonishing I know ! :)

Each point of femr2's alleged Savitzky-Golay acceleration curve was influenced by points within a window that was about twice as wide as femr2 has claimed.
Nope. You're flapping around in the dark still. Claimed ? Moi ? lol. What happened to me not providing you with details ?
 
Last edited:
...following which you've made about 10 posts within this thread and said nothing about the information provided to you,
Yes, I'm reading and following, and can only post short responses so far. It takes quite some time for me to make the experiments I want to do, and quite some time to elaborate a complete response. I have quite some topics pending to respond, and you keep piling them up. Not sure if I will reply to all; I probably won't.


What were you basing your understanding upon beforehand ?
The explanations posted in the forum. Which matched quite well the content of the paper, by the way, though the paper of course was richer in detail (such as the interesting part I quoted above).


1 sample width ! Ew :(
Yes, and it discards 1 sample at each end, which would be more if the window was wider.

By the way, a newer version of QtiPlot than the one I have also integrates data. If I can install it, I hope it will help me post an integral of your curve, which will hopefully be a much better approximation to the actual data your curve is a derivative of, than the one you provided. That's one of the things I haven't gotten to yet.
 
Yes, I'm reading and following, and can only post short responses so far. It takes quite some time for me to make the experiments I want to do, and quite some time to elaborate a complete response. I have quite some topics pending to respond, and you keep piling them up.
Only at your request/prompt.

Not sure if I will reply to all; I probably won't.
You rejected early motion assertions...I've reposted the information assertions are based upon...

As long as you answer at some point. Not a fan of totally wasting my time.

Strikes me that you'll not want to "agree" that early motion is present, as that would mean "agreeing" with femr2, which obviously would be a no-no...though quite WHY you'd want to reject the presence of early motion is beyond me.

The explanations posted in the forum. Which matched quite well the content of the paper, by the way, though the paper of course was richer in detail (such as the interesting part I quoted above).
Can't say I understand why you're spending so much time delving into low level SG behaviours...except of course to complain about 1% of my graphs...?

Yes, and it discards 1 sample at each end, which would be more if the window was wider.
Of course, but adjacent sample window is totally unsuited for the Dan Rather data. See page 1 of this 68 page thread.

By the way, a newer version of QtiPlot than the one I have also integrates data.
You don't have the data.

If I can install it
Why not just integrate with excel or similar ? Not rocket science.

I hope it will help me post an integral of your curve
Going to be rather difficult without the numerical data.

which will hopefully be a much better approximation to the actual data your curve is a derivative of, than the one you provided.
Of course it wouldn't...what do you think I've "provided" ?

That's one of the things I haven't gotten to yet.
Again, you're going to have problems integrating the data without having the data.

Quite why you're STILL focussing upon the "old" SG graph is "beyond me" :rolleyes:. You have a 1-pass SG curve which you have no basis to "complain" about...but are ignoring.

Why ?
 

ISF - Join now!

Every member here is approved by hand. No bots, no spam, just people who care about evidence and honest debate.

Membership is free!

Create your free account

Back
Top Bottom