The 9/11 Forum

einsteen wrote:I thought I posted this afternon, but guess I pressed the wrong button. I had no idea that you guys are hiding here, a lot of info here.

A big smile came to my face when I opened the page and saw you'd posted. And what a post! The video is awesome, it's amazing you can go through the whole process from front to back in 5 minutes - even with edits. Very clear and easy to follow. I appreciate your input on this thread; if it weren't for the first smearograms you posted at physorg, I'm quite sure this thread wouldn't exist.

Cool graphs, it is not strange that the blue and green one are almost the same, they are made using 2 videos from the same Naudet DVD, DC had some problems with ripping them, the only difference is the area at which the smearograms are taken, the offset in both directions should only be a translation in the plot.

Yes, they match up pretty well. The difference between your two curves could indeed be due to location. This is something we probably need to sort out. There's no doubt there should be difference between different horizontal positions on the roofline.

Another possibility is if there's a difference in the colors between the two copies, the decrease in color depth may result in a slightly different assignment in the border region between the two colors ('x' and 'not x'). I think I saw in your video the 2-color mask is based on nearest color. More on that in a bit.

Indeed the tower moves also a little bit to the right during the collapse and therefore its track is longer and therefore it should be a small correction if it didn't tilt, it also tilts in the other dimension. the NIST uses some professional techniques in their appendix, but I couldn't really find (using the search function...) how they estimated the collapse itself ?

I haven't seen anything in that regard either, but I confess to just skimming except for examining the details Dr. G has referenced. The parts that I have read are very impressive.

OneWhiteEye, your red line is interesting, is that the exponential one that dr. G fitted...

It's data taken from the same run, but not the same set. The primary data, which wanders a few pixels before drop, is based on the total area and intensity of a large feature. The red line is some auxiliary data, pixel-resolution, representing the bottom coordinate of the bounding box of the feature. It was the most stable of the lot, and not terribly so.

Here as promised the primitive method to create smearograms, OneWhiteEye gave me some ruby scripts but after I restored an O/S image I never used it again, sorry OWE...you're the man of that!

http://www.megaupload.com/?d=JNMLLCLW

Great video. The scripts have evolved a lot. But the setup is a hassle, no doubt. I bought a new quad-core a few months back and I still haven't put all the libraries on to do this stuff.

OWE/Einsteen:

Sorry to say I can't run Megaupload on my old Windows 98 machine.

Oh well, c'est la vie!

Dr. G wrote:OWE:

I really can't comment on how NIST arrived at its t(0)......

Hmm... it will be determined by examination, but this is another omission of critical information required for even the most basic validation of their work. I believe it is central.

I would say this however: Although NIST claim that roofline motion is detectable 6.9 seconds after the East Penthouse collapsed, you wonder why the set of images in Figure 5-205 start at t = 7.3 seconds. Interestingly, the next image, (at 7.8 seconds), still shows no real downward motion of the roof at the NE and NW corners, but we do see the start of a roofline bowing motion near the center of the north face.

This is curious and, again, difficult to resolve without just a few simple specifics that seem to be missing. I won't rule out sensitive motion detection on their part, which might not show up in the low-res figure they provide.

The plot you posted showing your, (near-center of roofline motion), data vs. Einsteen's (NW corner of roofline motion) data is very revealing in this regard. I would say the center motion shows a t(0) ~ 7.3 seconds, while the NW corner motion has a t(0) ~ 7.8 seconds - or 1/2 second later.

I would also say that the bowing which started at 7.3 seconds eventually pulled the center of the north face roofline down by as much as 2 meters. This behavior is nicely reflected in your curve vs. Einsteen's curve, which shows that after t = 8.3 seconds, the drop of the center of the roof was always about 2 meters ahead of the NW corner drop.

I'm glad that the pieces fit together pretty well. More pieces coming.

Dr. G wrote:OWE/Einsteen:

Sorry to say I can't run Megaupload on my old Windows 98 machine.

Oh well, c'est la vie!

Keep that machine! There's a lot of nasty stuff that won't run on it, either. It's an excellent grandfather technology. My old 98 box is dead, unfortunately.

There are ways around this technical difficulty, if only email, for any video you must examine. einsteen's video shows what's happening on-screen as he goes through each step, using the various programs and a script to go from frames of video, to smear image, to monchrome mask, to numbers, to curve fit - all in 5 minutes of video.

OWE:

Well, as an "old-geezer" I naturally have a lot of old "stuff":

My car is a '95 Ford Taurus; my CD player is a Rotel model purchased in 1992. I still watch a Sony TV purchased in 1986; and listen to records on my 1984 Thorens turntable. My camera is a Canon AE 1 purchased in 1976 and my watch is an Omega Seamaster from 1968.... They just don't build stuff like this anymore,..... so sure, I'll stick with my Windows '98 computer for as long as I can!

Dr. G wrote:OWE:

Well, as an "old-geezer" I naturally have a lot of old "stuff":

My car is a '95 Ford Taurus; my CD player is a Rotel model purchased in 1992. I still watch a Sony TV purchased in 1986; and listen to records on my 1984 Thorens turntable. My camera is a Canon AE 1 purchased in 1976 and my watch is an Omega Seamaster from 1968.... They just don't build stuff like this anymore,..... so sure, I'll stick with my Windows '98 computer for as long as I can!

Hey, that's cool (except maybe the Taurus?). They don't build stuff like they used to. My grandmother's gas stove (with oven) weighed a ton and lasted decades without anything ever malfunctioning. Same with the washing machine, which had a wringer. And, except for a belt or things like that, the refrigerator. They're all probably still in use in some yuppie's retro kitchen somewhere. The stuff I've had, on the other hand, is junk - but not as easy to recycle when it fails prematurely and permanently from some weak link.

Malware does not always run on Windows 98, not anymore. It's easier (and some cases just plain possible) to have some control over the machine and software you allegedly own. The successors make it all the more easier for someone else to own and operate the system you paid for.

OWE, yes it is fun a video like that...I had to trim and crop the video parts a lot to avoid privacy things from the desktop...family photos etc.

Dr. G, will rapidshare work ?

http://rapidshare.com/files/149221658/W ... o.avi.html

ps. My car is even older but will soon be replaced by a Toyota Prius, no I'm no environmentalist but when I lease that one it differs a lot of tax. And dr.G. the only computer games I play (2 times a year I think) are....old MSX games on a emulator.

I don't know if 3D renderings of image intensity is as helpful for others as it is for me. It lets me distinguish intensities better than real brightness does. This is a section of the first frame of the Naudet video:




With intensity used as 'elevation' in a 3D rendering:


Different view, with a filled-in surface. The corner of the building is the low dark spot. Blocky compression artifact is evident in the blue sky 'plateau' as strips of slightly different elevations. The 'cliffs' are the target:


This is a side-view rendering that uses isocontours, lines are drawn at intensity increments. You can see there is a substantial region of gradient between sky and north face, and between the west and north faces.


Making a monochrome mask involves choosing which pixels are light and dark based on some rule, usually a threshold value. If the color depth is decreased to monochrome, and no value is specified, presumably those intensities greater than 127 are assigned white. In this case, the sky is about 200 and the building around 85, so the mask would be created with an edge cutting through a short way up the cliff.

Now, imagine the same section from either a different video, or the same video after filtering or some transformation - unless the overall average intensity levels happen to be the same, the edge will be placed in a different location. Practically speaking, this wouldn't count for much with close copies, a pixel here and there. Nonetheless, it is a variable that must be controlled, as it could easily show a 2 pixel offset with a difference in compression algorithms in the section shown above.

I took some data on the NW corner using the same 2D method used on the dark rectangle, but applied as a 1D method. There were separate runs for horizontal and vertical dimension using only corner and roofline edges, respectively. Naturally, I was interested in how my horizontal data compared to horizontal motion published by NIST in Appendix C of NCSTAR 1-9. NIST uses inches for units of displacement on their chart, so I will, too. At this horizontal location, 13 stories is 299 pixels => 0.172m/px or 6.78 in/px.

This is what my horizontal data looks like:




This is what the NIST data looks like:

larger: http://i35.tinypic.com/2mwc9bq.png



This is what the two look like overlaid:




Both show precisely (as you'll ever get) a maximum of 14 inches of travel and a period of about 4 seconds. The curves diverge at the onset of collapse, but the match is incredible until that point. Considering the datasets originated with different methods, they validate each other quite well.

I have some similar vertical displacement data to publish shortly. Unlike the horizontal, which is clearly validated over most of the period, I have to manually validate the vertical - which will take some time.

thank you so much for this topic, this is very interesting. Amazing work, keep it up

Dictator Cheney wrote:thank you so much for this topic, this is very interesting. Amazing work, keep it up

You're welcome, and thank you.

Yes, OneWhiteEye, who needs the NIST when we have two of our very own video analysis geniuses, (or is that genii) - you and Einsteen - right here on this forum!!!!

Way to go.... and looking forward to seeing a calibrated curve for the vertical oscillation data!

By the way, this makes me wonder how much of this type of video analysis NIST has already done but has not made public.....

Cool, OWE, that's a match! I still have to check out your red line and would like to be able to reproduce it with the childish smeorogram (word invited by dr. G...) method. What is the interpretation of those nice 3d plots ? It's late...

I still believe that the smeorograms could provide more data than one would expect at first because they are really smooth, noise is the limit but if there are more points one could effectively get a bigger accuracy, if something is at a point between [0,1] then if it found more at position 0 than position 1 the chance is higher that the real location is at <0.5 and vice versa, if there are enough points and the graphs are smooth it really says something about the motion. The real motion must also be smooth because of the big mass. On the one hand more points are needed and need to be averaged but on the other hand if you take the whole roof of the building then the average says nothing about the details. A simple extension of the smearogram idea is taking 5xn images....



this could be averaged to get a good right corner motion, simply taking the average of 5 points and so on.

Dr. G wrote:Yes, OneWhiteEye, who needs the NIST when we have two of our very own video analysis geniuses, (or is that genii) - you and Einsteen - right here on this forum!!!!

I favor 'genii' and there are three - before einsteen you were the only game in town.

Way to go.... and looking forward to seeing a calibrated curve for the vertical oscillation data!

Unfortunately, the vertical data did not pass muster. It went off target towards the end. There's no way I could have done the horizontal curve above by hand, yet I could have done a lot better manually with the vertical data. It happens; just have to set a better threshold for the later frames. Probably will not be able to get back to it for a day or two. Stupid job gets in the way.

By the way, this makes me wonder how much of this type of video analysis NIST has already done but has not made public.....

Hmmm. Me, too.

einsteen wrote:I still have to check out your red line and would like to be able to reproduce it with the childish smeorogram (word invited by dr. G...) method.

Ahh, yes, smear-o-gram. Doesn't sound serious but it stuck... I called them 'einsteen images' for a stretch, but that's a little weird. The one you did of the second impact was really cool.

It's the best way to produce a graph of motion immediately from video. In fact, it's a great end product in itself. It's the only honest end product since everything else adds subjectivity and error. The final numeric output, while only pixel-resolution, is generally very smooth compared to what I'm doing despite having the same time resolution. And, as you say, multiple snapshots can be combined to increase the resolution.

The red line is the lowest point on the large dark rectangle, top center of of the north face. This should be the lower left corner of the rect. I don't know how good that data is; it's pretty jumpy compared to your curves.

What is the interpretation of those nice 3d plots ? It's late...

The third dimension, height, corresponds to the luminance of the pixel. The brighter a pixel is, the higher the corresponding point on the surface (unless inverted). The floor of the graph, by default, is 0 and the ceiling 255, giving a spatial interpretation of intensity. Intensity gradients appear as slopes. Reducing an image to two colors based on a threshold is seen here as cutting through the surface with a horizontal plane, usually at half-height, then setting the points below the surface to black and those above to white.
















hahahaha.... hope you weren't eating.

I still believe that the smeorograms could provide more data than one would expect at first because they are really smooth, noise is the limit but if there are more points one could effectively get a bigger accuracy, if something is at a point between [0,1] then if it found more at position 0 than position 1 the chance is higher that the real location is at <0.5 and vice versa, if there are enough points and the graphs are smooth it really says something about the motion. The real motion must also be smooth because of the big mass. On the one hand more points are needed and need to be averaged but on the other hand if you take the whole roof of the building then the average says nothing about the details. A simple extension of the smearogram idea is taking 5xn images....

---image---

this could be averaged to get a good right corner motion, simply taking the average of 5 points and so on.

Absolutely. In fact, you can also average across a range of thresholds used to create the mask. That will pick up the contour from intensity gradient variations in the curve and be very close to what I'm doing.