The 9/11 Forum

RGB channels:





The bump starting around frame 880 is smoke passing by. The displacement scale (y-axis) is in pixels. Note the gridlines are spaced 0.1 pixel (~2.5cm) apart.

All channels averaged:





Comparison with old data on top center dish, frames 800-920:

But it's hard to take these fine motions too seriously....

After a run with the new method on both the top center and top right dishes, the separation over time is:



They drift about 0.4 pixels closer to each other over frames 800-880. I've no explanation for this, other than changes in illumination over time, from the smoke plume. Casts some doubt on anything based on fine motion from the earlier data.

Average of four dishes (TC,TR,BC,BR); looks like that fine motion is real. Again. Who knows? I'll find out.

Unfortunately, none of the dishes can be followed past frame 990. Looking for suitable 'blob' targets higher up the antenna, here is one (or two)...

Acquiring one of many stationary targets in the scene, frames 800 - 1020, vertical displacement. The vertical scale spans 0.3 pixels. Stays within a band of about +/- 0.05 pixels, except for a few places.

A very accurate measurement of the horizontal motion of the camera.

0.3 pixel span on the y axis.

Some notes:

My frame numbers are 10 off from what is normally seen, including my first rip of an internet video. The subsequent rip straight from DVD had the frame number offset but I published data with those numbers after telling interested parties (party) that new frame numbering was in effect. The last rip for DVD had the right frame numbers but I renumbered to match the previous rip 'to avoid counfusion'. Yikes. I'm thinking of renumbering this set to correct frame numbers.

Initiation is around frame 907 in this latest series, 917 otherwise and possibly in the future.

The displacement graphs end around initiation. That's because progression is B-O-R-I-N-G from a data acquisition perspective. The stationary targets, for measuring camera wobble, run from 800-1020 to catch most of the period that's been examined to this point.

I was happy to put these graphs up because they represent the first fruits of many hours of labor in study, programming, and what have you. The process is much more sophisticated than days gone by. The data taken above uses adaptive thresholding (a must) and some blob classification, both new for me. The RGB channels are recorded individually for variable post-processing methodology, whereas they were simply averaged before.

It does just scratch the surface, though; by building my own 'SynthEyes' I get optimal control over every jot and tittle of the process. Supervised and unsupervised learning can be exploited to any degree. What I learn about the process can be rolled into code and repetitive 'intelligent' operations can be automated and parameterized.

Metadata, too. It can know when data is suspect or bad without even examining the data itself. This plot of a bad tracking has markers where size and color indicate 'quality'. Of course, the data is obviously bad but this sort of qualitative measure is useful when the distinction isn't clear ---



Frame viewing, cropping, filtering, feature extraction, tracking, and graphing of any sort all rolled into one program!

Hey, I might add curve fitting and parameter estimation someday, too!

Edit: not to mention inline parametric perspective correction with error bands and automatic correlation of datasets spanning various intervals.

Piece of cake now.

OneWhiteEye wrote:Piece of cake now.

Wassat ?

And.....release !

femr2 wrote:Wassat ?

The east side of the north roofline of WTC1 from the Sauret video. Frames and pixels. Incomplete, of course. Unvalidated and without process description.*

And.....release !

The numeric data? Very soon. I want to get most of frames 800-964 first. As odd as it may seem, this thread is for talking about how to acquire data as opposed to releasing data. Actual datasets come up, naturally, and sometimes people do things with them because they're the most accurate and/or complete sets they've run across. And, yes, I've promoted them as such, with the usual caveats about this or that type of error. Hopefully, something definitive on a feature will be possible soon, but it's been only an exercise to this point. All datasets have been provisional.

The process of getting there with measurement is sometimes very interesting to me, sometimes not at all. For most, I'm sure, it's 'not at all'.

The flipside, though, is always interesting to me - what people do with the data. I only know of five people that have done anything (including simply play) with my data. Small audience, but a rather concentrated one, if you ask me. Did you have any specific purpose in mind? I know what I'd do with the data, but what do others want to do?

*edit: and probably still not as accurate as it could be

It is still useful to look at curves, no doubt.


The roofline curve together with two of the dishes:




The vertical pixel separation between roofline and a dish over the 25 frames in common:

The NE corner of the tower:



There it is. The roofline. Not all by itself, but very distinct. Now the modified k-means clustering technique must be completed to select the ONE mostly horizontal line. The processing may be pre-optimized by applying a Sobel horizontal edge detection on top of this.

Got it! No clustering required. It's light aqua now:



Who cares that it's the whole corner and not just the roofline? It's better. Wonder how well this works around frame 940?

Eh, clustering will be better. The heuristic is the roofline angle.